Radar working mode identification method based on prior matching and inter-domain alignment

Abstract

The invention discloses a radar working mode identification method based on prior matching and domain distribution alignment. Firstly, radar pulse description word data are preprocessed, abnormal values, noise and the like are removed, and the quality of the data is improved. Then, an encoder is used for extracting potential features from radar pulse description word data detected by a source domain and a target domain, supervised training is carried out in the source domain through the potential features of the source domain data and corresponding labels, and prior distribution of the source domain features and prior distribution of the target domain features are restrained to be standard normal distribution by constructing prior matching regularization items. A distance based on distribution characteristic functions is used as a measure to realize explicit calculation of domain distribution alignment. And finally, constructing a loss function by using a prior matching regularization term and a distribution characteristic function distance, and carrying out unsupervised training on data of the target domain to improve the performance of the radar working mode recognition model in the target domain.

Description

Technical Field

[0001] The invention relates to a radar working mode recognition method and belongs to the field of radar working mode recognition. Background Art

[0002] Radar operating mode recognition is a technology involving signal processing, pattern recognition and machine learning. Its purpose is to infer the specific operating mode of the radar system based on the characteristics of the received radar signal. Operating mode recognition of the received radar signal can help analyze the other party's behavior intentions and assess the situation.

[0003] The recognition methods of radar working modes mainly include traditional feature-based methods, machine learning-based methods, deep learning methods, etc. Since deep learning models can automatically learn high-level abstract representations of input data, they perform well in radar working mode recognition. However, due to the differences in the features of pulse descriptors of different working modes of radars in different environments, the model trained in the source domain will have performance degradation in the target domain. Summary of the invention

[0004] Aiming at the problem that the radar working mode recognition deep learning model trained in the source domain has poor adaptability in the target domain, the present invention proposes a radar working mode recognition method based on prior matching and inter-domain alignment, which can explicitly calculate domain differences, aiming to improve the adaptability of the radar working mode recognition model in the target domain, thereby improving the accuracy of radar working mode recognition and providing strong support for one's own judgment and decision-making.

[0005] The present invention provides a radar working mode recognition method based on prior matching and inter-domain alignment, comprising the following steps:

[0006] Step (1) preprocessing the data to remove outliers and noise, etc., to improve the quality of the data;

[0007] Step (2) using an encoder to extract latent features from radar pulse descriptor data detected in the source domain and the target domain respectively, and using the latent features of the source domain data and the corresponding labels to perform supervised training in the source domain;

[0008] Step (3) constrains the prior distribution of source domain and target domain features to be a standard normal distribution by constructing a prior matching regularization term;

[0009] Step (4) uses a distance between distribution feature functions as a metric to achieve explicit calculation of domain distribution alignment;

[0010] Step (5) combines the prior matching regularization term and the distribution characteristic function distance to perform unsupervised training on the target domain data to improve the performance of the radar working mode recognition model in the target domain.

[0011] The beneficial effects achieved by the present invention are as follows: The present invention solves the problem that the performance of the radar working mode recognition model trained in the source domain degrades in the target domain. The prior matching regularization term is introduced into the prior distribution matching of the source domain and target domain features, which can better constrain the space of potential features and make the learned feature distribution have better generalization ability. A new metric based on the distribution feature function distance is proposed, which converts the difference between the feature distributions into the frequency domain for calculation, and can explicitly calculate the difference between the source domain and the target domain. The target domain data is unsupervisedly trained through the prior matching regularization term and the distribution feature function distance to improve the performance of the radar working mode recognition model in the target domain. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 This is a flow chart of an embodiment of the present invention DETAILED DESCRIPTION

[0013] The present invention proposes a radar working mode recognition method based on prior matching and inter-domain alignment, uses an encoder to extract the potential features of radar pulse descriptor data in the source domain and the target domain, and constrains the feature distribution of the source domain and the target domain to be a normal distribution by constructing a regularization term, which can better constrain the space of potential features and make the learned feature distribution have better generalization ability. A distance based on distribution feature function is proposed as a metric to explicitly calculate regional differences in regional adaptation, and a distribution alignment total loss function is defined to perform unsupervised training on the data in the target domain to improve the adaptability of the model.

[0014] The present invention will be further explained below in conjunction with the accompanying drawings and embodiments.

[0015] The schematic diagram of the embodiment of the present invention is as follows Figure 1 As shown. The radar working mode recognition method based on explicit distribution alignment of domain difference measurement proposed by the present invention, the encoders used in the preferred embodiment are all existing residual networks. The preferred embodiment process of the present invention includes:

[0016] (1) Preprocess the data to remove outliers and noise, etc., to improve the quality of the data. The preferred implementation process includes:

[0017] is the radar pulse description word data N received in the source domain S The correct radar working mode classification is From the target domain, its radar working mode classification unknown.

[0018] The box plot method is used to remove outliers in the data, including noise and interference: outliers are defined as values ​​less than the lower limit or greater than the upper limit, the lower limit is δ1=Q1-1.5*IQR, and the upper limit is δ2=Q2+1.5*IQR, where Q1 is the lower quartile, Q2 is the upper quartile, IQR=Q2-Q1 is the interquartile range, and 1.5 times the IQR is an empirical value.

[0019] The encoder is used to extract latent features from the source domain data and the target domain data respectively, and supervised training is performed in the source domain using the latent features of the source domain data and the corresponding labels. The preferred implementation process includes:

[0020] The latent feature z is obtained from the input data X through the encoder module s or X T The encoder module is a mapping F that converts the source domain data X s and target domain data X T Transformed into feature z in the latent feature space S and z T ,Right now:

[0021] z S / T =F(X S / T ) (1)

[0022] When the original domain data X s and target domain data X T When passed to the encoder, the encoder will output z S and z T The parameter is the mean u S / T and variance∑ S / T These parameters are used to define the posterior distribution q S (z|x) and q T (z|x), the posterior distribution can be expressed as q S / T (z|x)=N(u S / T ,Σ S / T ) where N represents the normal distribution By using the posterior distribution q S (z|x) and q T Random sampling in (z|x) can obtain the potential feature z of the source domain S and the latent feature z of the target domain T .

[0023] By S Decoding to get the prediction of source domain classification The learning of the source domain radar working mode recognition model is supervised by the following formula:

[0024]

[0025] Where L CE Denotes the cross entropy loss represents the classification loss, n S Represents the number of original domain samples.

[0026] (2) By constructing a priori matching regularization term, the prior distribution of source domain and target domain features are constrained to be a standard normal distribution. The preferred implementation includes:

[0027] Prior matching regularization loss L prior Given by:

[0028]

[0029] where KL(·||·) represents the KL divergence between two distributions, E(·) represents the expected value, and 0∈R n×n is a vector with all elements equal to zero. The total loss function L of the proposed distribution alignment framework is as follows:

[0030] L=α1L prior +α2L dd (4)

[0031] Where L dd is the explicit distribution alignment loss function, and α1 and α2 are parameters that control the trade-off between the two loss functions.

[0032] (3) Using a distance between distribution feature functions as a metric to achieve explicit calculation of domain distribution alignment, the preferred implementation process includes:

[0033] Define the distribution characteristic function:

[0034]

[0035] in, is the characteristic function of the distribution function p(z), i represents the imaginary unit, yes The transpose of , where E(·) represents the expected value, and when calculating formula (5) The Fourier transform of To do the calculation:

[0036]

[0037] and in is the frequency, and F(·) represents the Fourier transform of the function. The distribution feature function distance is used to measure the domain difference and realize the display calculation of domain distribution alignment. The l2 norm is used to calculate the distribution feature function distance, and the loss function is given by the following formula:

[0038]

[0039] Among them, ||·|| 2 is the square modulus of the complex number, Determines the size of the integration region.

[0040] Then use batch samples to estimate the distribution characteristic function distance:

[0041]

[0042] in for z Sk Yes S The kth element of Tk The same is true; M is the number of samples.

[0043] When z S and z T When the dimension n is large, the slice distribution characteristic function distance is defined as:

[0044]

[0045] where the projection parameter θ is Medium unit sphere S n-1 The uniform measure on According to experience, we select the same number of fixed projection parameters {θ i |i=1,…,n}, where θ i ∈R n is a standard unit vector whose i-th element is 1. In addition, the first-order statistical information, namely the mean, is used to constrain the distribution alignment. S and z T The Euclidean distance of the expected value defines the average loss L mean ,as follows:

[0046]

[0047] Combining the slice distribution feature function distance with the average loss, we get the distribution feature function distance loss, i.e., the explicit domain difference loss L explicit :

[0048] L explicit =β1L SCF +β2L mean (10)

[0049] Among them, β1 and β2 are trade-off parameters.

[0050] (4) The prior matching regularization term and the distribution characteristic function distance are combined to perform unsupervised training on the target domain data to improve the performance of the radar working mode recognition model. The implementation process is as follows:

[0051] Substituting equation (10) into (4), the total loss function of distribution alignment is obtained as follows:

[0052] L=α1L prior +α2(β1L SCF +β2L mean ) (11)

[0053] By using formula (11) to perform unsupervised domain distribution alignment training, we can obtain the radar working mode recognition model in the target domain and realize the radar working mode recognition in the target domain.

Claims

1. A radar working mode recognition method based on prior matching and inter-domain alignment, characterized in that: Step (1): preprocess the data to remove outliers and noise; Step (2) using an encoder to extract latent features from radar pulse descriptor data detected in the source domain and the target domain respectively, and using the latent features of the source domain data and the corresponding labels to perform supervised training in the source domain; Step (3) constrains the prior distribution of source domain and target domain features to be a standard normal distribution by constructing a prior matching regularization term; Step (4) uses a distance between distribution feature functions as a metric to achieve explicit calculation of domain distribution alignment; Step (5) combines the prior matching regularization term and the distribution characteristic function distance to perform unsupervised training on the target domain data.

2. The radar working mode recognition method based on prior matching and inter-domain alignment according to claim 1 is characterized in that: The step (2) further comprises: From the input data X through the encoder module s or X T The encoder module is represented as a mapping F, which transforms the source domain data X s and target domain data X T Transformed into feature z in the latent feature space S and z T ,Right now: z S / T =F(X S / T ) (1) When the original domain data X s and target domain data X T When passed to the encoder, the encoder outputs z S and z T The parameter is the mean u S / T and variance∑ S / T , used to define the posterior distribution q S (z|x) and q T (z|x), the posterior distribution is denoted by q S / T (z|x)=N(u S / T ,Σ S / T ), where N represents the normal distribution By using the posterior distribution q S (z|x) and q T Random sampling in (z|x) to obtain the potential feature z of the source domain S and the latent feature z of the target domain T ; By S Decoding to get the prediction of source domain classification The learning of the source domain radar working mode recognition model is supervised by the following formula: Where L CE Denotes the cross entropy loss represents the classification loss, n S Represents the number of original domain samples.

3. The radar working mode recognition method based on prior matching and inter-domain alignment according to claim 1 is characterized in that: The step (3) further comprises: Prior matching regularization loss L prior Given by: where KL(·||·) represents the KL divergence between two distributions, 0∈R n×n is a vector with all elements equal to zero; the total loss function of the proposed distribution alignment framework is as follows: L=α1L prior +α2L dd (4) Where L dd is the explicit distribution alignment loss function, and α1 and α2 are parameters that control the trade-off between the two loss functions.

4. The radar working mode recognition method based on prior matching and inter-domain alignment according to claim 1 is characterized in that: The step (4) further comprises: Define the distribution characteristic function: in, is the characteristic function of the distribution function p(z), i represents the imaginary unit, yes The transpose of , where E(·) represents the expected value, and when calculating formula (5) The Fourier transform of conduct: and: in is the frequency, F(·) represents the Fourier transform of the function; the distribution feature function distance is used to measure the domain difference, showing that the calculation domain distribution is aligned; the l2 norm is used to calculate the distribution feature function distance, and the loss function is given by the following formula: Among them, ||·|| 2 is the square modulus of the complex number, Determine the size of the integration region; Then use the mini-batch samples to estimate the distribution characteristic function distance: in for z Sk Yes S The kth element of Tk The same is true; M is the number of samples; When z S and z T When n is large, the slice distribution characteristic function distance is defined as: where the projection parameter θ is Medium unit sphere S n-1 The uniform measure on According to experience, we select the same number of fixed projection parameters {θ i |i=1,…,n}, where θ i ∈R n is a standard unit vector whose i-th element is 1; in addition, the first-order statistical information, namely the mean, is used to constrain the distribution alignment; z S and z T The Euclidean distance of the expected value defines the average loss L mean ,as follows: Combining the slice distribution feature function distance with the average loss, we get the distribution feature function distance loss, i.e., the explicit domain difference loss L explicit : L explicit =β1L SCF +β2L mean (10) Among them, β1 and β2 are trade-off parameters.

5. The radar working mode recognition method based on prior matching and inter-domain alignment according to claim 1 is characterized in that: The step (5) further comprises: The total loss function L for distribution alignment is as follows: L=α1L prior +α2(β1L SCF +β2L mean ) (11) Formula (11) is used to perform domain distribution alignment training and obtain the radar working mode recognition model in the target domain.

Images