Intelligent radar hrrp target recognition method and system

Abstract

The application discloses an intelligent radar HRRP target identification method and system, relates to the technical field of target identification, and comprises the following steps: in the CNN-BTEN network, a convolution module is used for respectively extracting features of multi-dimensional range images in a radar high-resolution range sample image; a position coding module is used for coding positions of a first feature map sequence by using a trigonometric function; a BCTE module is used for extracting features of a second feature map sequence based on a Bayesian linear regression and a Transformer encoder; a classification module is used for calculating probabilities of various ship types in a third feature map to realize ship target classification; a training sample set is used for training the CNN-BTEN network to obtain a target ship identification model; and a radar high-resolution range image to be identified is input into the target ship identification model to obtain a ship type in the radar high-resolution range image to be identified. The application realizes fast and accurate identification of ships.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of target identification, in particular to an intelligent radar HRRP target identification method and system based on a CNN-Bayesian-Transformer neural network. BACKGROUND

[0002] At present, a traditional identification method is mainly based on shallow structural features of radar data, and part of information is inevitably lost, so that the final HRRP (High Resolution Range Profile) target identification precision is limited. Due to the excellent nonlinear feature extraction capability of deep learning, the deep learning has become a research hotspot in various fields, and has been widely applied to target detection, identification and classification in synthetic aperture radar, inverse synthetic aperture radar and optical and remote sensing images. Compared with the traditional identification method, the deep learning algorithm can autonomously learn higher-level structural information and complete classification and identification of different targets. The classification and identification of radar based on target HRRP information can also be realized through the deep learning algorithm, and a lot of research results have been achieved in this field.

[0003] There is no typical public database for HRRP target identification based on deep learning at present, and the databases for HRRP target identification of different teams and researchers are mostly from model simulation calculation and darkroom measurement of the teams. However, different literature researches show that the HRRP target identification accuracy based on deep learning is higher than that of the traditional classification method, and typical application methods include CNN, stacked auto encoder (SAE), recurrent neural network (RNN), deep belief networks (DBN) and the like. Among them, the CNN can combine feature extraction and feature-based classification and identification, and can extract higher-level feature information from the HRRP signal through convolution calculation and mine potential data correlation information. At present, the CNN has been effectively applied to the RATA field.

[0004] However, due to the limited receptive field, it is difficult to learn the global features of HRRP using a CNN model. Although CNN can capture important physical structure features of HRRP data, the convolution operation of CNN can only capture local information each time, and can only enhance the receptive field through multiple stacking of layers to obtain the connection between larger regions, so it is difficult to capture global information. Therefore, a shallow CNN network cannot capture long-distance features. The failure of CNN to capture the time series correlation across the distance units of HRRP will result in the loss of valuable information reflecting the physical structure features of the target. In addition, saturation problems may occur, and as the number of network layers increases, the recognition accuracy will decrease.

[0005] The recurrent neural network (RNN) model in deep learning has shown excellent sequence modeling capability in natural language processing, machine translation and speech recognition. In the field of RATR, RNN is used to extract long-term time structure embedded in HRRP sequence. However, the dimension and length of the input sequence in the RNN model are coupled and cannot be adjusted independently, which reduces the flexibility of the model. In addition, as the sequence length increases, the long-range correlation will be severely weakened, thereby limiting the application of deep models in HRRP target recognition. SUMMARY

[0006] The purpose of the present application is to provide an intelligent radar HRRP target recognition method and system, which can realize fast and accurate identification of ships.

[0007] To achieve the above purpose, the present application provides the following scheme:

[0008] In a first aspect, the present application provides an intelligent radar HRRP target recognition method, comprising:

[0009] obtaining a training sample set; each training sample in the training sample set comprises a radar high-resolution range sample image and a ship type corresponding to the radar high-resolution range sample image;

[0010] constructing a CNN-BTEN network; the CNN-BTEN network comprises a convolution module, a position encoding module, a BCTE module and a classification module; the convolution module is used for feature extraction on multi-dimensional distance images in the radar high-resolution range sample image to obtain a first feature map sequence; the position encoding module is used for position encoding on the first feature map sequence by using a trigonometric function to obtain a second feature map sequence; the BCTE module is used for feature extraction on the second feature map sequence based on Bayesian linear regression and a Transformer encoder to obtain a third feature map; and the classification module is used for calculating the probability of each ship type in the third feature map to realize ship target classification;

[0011] The CNN-BTEN network is trained by using the training sample set to obtain an optimal CNN-BTEN network; the optimal CNN-BTEN network is a target ship identification model;

[0012] Obtain a radar high-resolution range profile to be identified;

[0013] Input the radar high-resolution range profile to be identified into the target ship identification model to obtain a ship type in the radar high-resolution range profile to be identified.

[0014] In a second aspect, the present application provides an intelligent radar HRRP target identification system, comprising:

[0015] A training sample determination subsystem is configured to obtain a training sample set; each training sample in the training sample set comprises a radar high-resolution range sample image and a ship type corresponding to the radar high-resolution range sample image;

[0016] A network construction subsystem is configured to construct a CNN-BTEN network; the CNN-BTEN network comprises a convolution module, a position coding module, a BCTE module and a classification module; the convolution module is configured to perform feature extraction on a multi-dimensional range image in the radar high-resolution range sample image to obtain a first feature map sequence; the position coding module is configured to perform position coding on the first feature map sequence by using a trigonometric function to obtain a second feature map sequence; the BCTE module is configured to perform feature extraction on the second feature map sequence based on a Bayesian linear regression and a Transformer encoder to obtain a third feature map; and the classification module is configured to calculate a probability of each ship type in the third feature map to realize ship target classification;

[0017] A network training subsystem is configured to train the CNN-BTEN network by using the training sample set to obtain an optimal CNN-BTEN network; the optimal CNN-BTEN network is a target ship identification model;

[0018] A to-be-identified image acquisition subsystem is configured to obtain a radar high-resolution range profile to be identified;

[0019] An image identification subsystem is configured to input the radar high-resolution range profile to be identified into the target ship identification model to obtain a ship type in the radar high-resolution range profile to be identified.

[0020] According to the embodiments of the present application, the following technical effects are achieved:

[0021] The application discloses an intelligent radar HRRP target identification method and system, and aims at the situation that the existing feature extraction method cannot achieve satisfactory results due to the complex data structure of radar high resolution range profile (HRRP), and the method comprises the following steps: constructing a CNN-BTEN (Convolutional Neural Networks-Bayes-TransEncoder Network) network based on a convolution module, a position coding module, a BCTE (Bayes Connection-Transformer Encoder) module and a classification module; extracting local spatial features of the HRRP by using the convolution module, and embedding position information to generate new time sequence features by using the position coding module, so that the early expression capability of the network model and the efficiency and flexibility of the HRRP modeling task can be greatly improved; adopting the BCTE module to combine a transformer encoder and a Bayesian linear regression, so that the dependency relationship between two distance units at any position is perfectly described by using the multi-head self-attention mechanism in the transformer encoder, and one-step local and global dependencies can be captured; meanwhile, the prior and posterior information of the Bayesian linear regression is used to effectively enhance the recognition probability. In addition, the BCTE module also has good parallelism; and finally, the classification module is used for classification. After the training sample set is used to train the above CNN-BTEN, a target ship identification model is obtained. When the target identification of the sea surface ship is performed, the radar high resolution range profile to be identified is input into the target ship identification model, so that the HRRP features can be stably and reliably extracted, and then the accurate ship type can be quickly obtained. BRIEF DESCRIPTION OF DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments. Obviously, the drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0023] Figure 1 It is a flowchart of the intelligent radar HRRP target identification method of the present application.

[0024] Figure 2 It is a HRRP schematic diagram of a certain type of target in the present application.

[0025] Figure 3 It is a HRRP schematic diagram of a certain type of target in the present application.

[0026] Figure 4 It is a schematic diagram of data enhancement on the HRRP in the present application.

[0027] Figure 5This is a schematic diagram of the CNN-BTEN network structure of the present invention;

[0028] Figure 6 This is a schematic diagram of the multinomial distribution probability of the six target types in this invention;

[0029] Figure 7 This is a comparison chart of the 10-epoch time consumption of the three models in a specific example of the present invention;

[0030] Figure 8 This is a schematic diagram of the intelligent radar HRRP target recognition system of the present invention. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] The purpose of this invention is to provide an intelligent radar HRRP target recognition method and system, which uses a novel deep learning model, CNN-BTEN network, based on one-dimensional convolutional neural networks (1D-CNN) and the Transformer architecture, to identify ships on the sea surface.

[0033] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0034] Example 1

[0035] like Figure 1 As shown, this embodiment provides a smart radar HRRP target recognition method, including:

[0036] Step 100: Obtain a training sample set; each training sample in the training sample set includes a radar high-resolution range sample image and the ship type corresponding to the radar high-resolution range sample image.

[0037] Step 100 specifically includes:

[0038] 1) Construct multiple types of ship models.

[0039] There are two methods to obtain the target echo signal: actual measurement method and theoretical calculation method. Since most ship targets are non-cooperative targets, it is very difficult to obtain HRRP by on-site measurement. Therefore, the application adopts 3D Max to establish six ship models for simulation.

[0040] 2) Simulate the plurality of types of ship models by using the FEKO software to obtain a plurality of initial radar high-resolution range profiles corresponding thereto; wherein the simulation parameters set include that the center frequency of the radar is 10 GHz, the bandwidth is 80 MHz, the number of frequency sampling points is 200, the calculated azimuth angle range is 0-360°, the interval is 1°, the grazing angle is 10°, and based thereon, the initial radar high-resolution range profile that can be obtained has 200 distance units, and the length corresponding to each distance unit is 1.875 m. A certain HRRP sample of one ship and the HRRP after amplitude normalization are respectively as shown in Figure 2 and Figure 3 In Figure 2 and Figure 3 , the horizontal axis and the vertical axis respectively represent the distance unit and the azimuth angle of the HRRP, and 360 HRRP data are obtained for each ship.

[0041] 3) For each of the initial radar high-resolution range profile, data enhancement processing or multiple addition of Gaussian white noise processing is performed to obtain a plurality of radar high-resolution range sample images. Specifically, in order to meet the requirement of sample data quantity during neural network training and prevent overfitting, the data set is expanded. The application selects two expansion methods:

[0042] A) Data enhancement processing. As shown in Figure 4 , when calculating the HRRP, the coordinate axis coincides with the ship center, and therefore the effective HRRP information is generally in the middle area. In the actual environment, when the radar detects the target, the echo signal may be incomplete or partially lost. Therefore, it is necessary to intercept the translation of the HRRP. Since the HRRP is one-dimensional data, only one-dimensional translation interception is applied. Then, by using the data enhancement technology, the HRRP signal is moved along the distance direction to expand the data set. Each time, the HRRP signal is moved by 20 distance units, moved left twice, and moved right twice. The deleted data is discarded, and the blank part is filled with zero-padding, i.e. the blank features are filled with 0, as shown in Figure 4 . By selecting overlapping but different HRRP samples, the number of samples is increased by 4 times.

[0043] B) Gaussian white noise addition processing. Random noise is added to the translated HRRP data. The data is added with 10 times of Gaussian white noise, and the data after adding the noise satisfies a certain signal-to-noise ratio.

[0044] 4) Determine the ship type of the ship model in each radar high-resolution range sample image; each radar high-resolution range sample image and the corresponding ship type of the radar high-resolution range sample image constitute a single training sample; a plurality of training samples constitute a training sample set.

[0045] Specifically, the training sample set is divided into a training set and a test set according to a ratio of 8:2 for each type of ship target, and the training set and the test set samples are 43200 and 10800 respectively. To ensure the authenticity and effectiveness of the network model classification and identification performance, the entire data set needs to be randomly shuffled and reorganized before being divided.

[0046] In a specific practical application, since the radar target object can be represented by a scattering point model, the HRRP contains the distribution of the target in the radar line-of-sight direction. However, the target often produces relative displacement in the radar line-of-sight direction, and the corresponding echo also changes relatively, so the data needs to be preprocessed. The preprocessing includes the following two steps:

[0047] 1) Normalize the amplitude of the HRRP. The normalized data of the nth HRRP amplitude is represented as where max(|x n |) represents the maximum value of the absolute value of all elements in the HRRP.

[0048] 2) Subtract the average value of the normalized HRRP data from the respective elements.

[0049] Through the preprocessed data, it is easier to find the minimum value in the subsequent gradient descent iteration process, so that the model converges faster.

[0050] Step 200, constructing a CNN-BTEN network; as shown in Figure 5 The CNN-BTEN network includes a convolution module, a position encoding module, a BCTE module and a classification module; the convolution module is used for feature extraction of the multi-dimensional range image in the radar high-resolution range sample image to obtain a first feature map sequence; the position encoding module is used for position encoding of the first feature map sequence using a trigonometric function to obtain a second feature map sequence; the BCTE module is used for feature extraction of the second feature map sequence based on Bayesian linear regression and a Transformer encoder to obtain a third feature map; the classification module is used for calculating the probability of each ship type in the third feature map to realize ship target classification.

[0051] (I) Convolution module

[0052] There are generally two methods for obtaining input sequence features: one is a time domain segmentation method adopted by a deep network such as an RNN or a long short-term memory (LSTM), which has the problems of information redundancy and dimension-length constraints; and the other is a direct segmentation method (patchify) adopted by a neural network such as a ViT, which is usually implemented by using a large convolution kernel and a large stride convolution, which violates the typical design of a convolution layer in a neural network and restricts the expression ability of the network at a fixed position in an early stage.

[0053] To solve the above problems, the application adopts a convolution operation to replace the above two methods to obtain sequence features as input. On the one hand, the extracted sequence features better preserve the local structure information of the HRRP; on the other hand, the dimension and the number of channels of the convolution kernel independently control the dimension and the sequence length of the features, respectively, decouples the two, and improves the representation ability of the entire network for HRRP data. In addition, the size and the stride of the convolution kernel are controllable, which can effectively avoid the constraint of the fixed position and further enhance the expression ability of the network in the early stage.

[0054] The so-called convolution operation in deep learning is also called cross-correlation operation: from left to right and from top to bottom in the input matrix, a part of the same size as the filter is taken, the values in each part are multiplied with the values in the filter, and the sum is taken, and finally the result forms a matrix, without flipping the kernel. The original one-dimensional range image is one-dimensional data, that is, a one-dimensional convolution is needed to filter and extract features. The larger the convolution kernel, the larger the "receptive field" of the convolution layer, that is, the larger the local connection area. In addition, the convolution kernel parameters in the same convolution layer are shared, which can reduce the parameters to be solved. The step is the number of units that the convolution kernel moves each time between two adjacent convolution operations of the convolution kernel. When there is no padding, the larger the step, the smaller the size of the output feature map. The padding is to make up for the loss of part of the value after the convolution of the input matrix and the convolution kernel, and the edge of the input matrix is "trimmed" (only part of the position points are detected at the edge, and a large amount of information at the boundary of the input matrix is lost). According to the padding effect, it can be divided into effective padding, same padding and full padding, that is, three modes of convolution: effective convolution, same convolution and full convolution. The padding can be performed on the boundary of the original matrix before the convolution operation, that is, the boundary of the matrix is filled with numerical values to increase the size of the matrix, which is usually filled with "0". The output feature matrix will be smaller, unchanged and larger than the input feature matrix, respectively. The commonly used padding methods are effective padding and same padding. The application selects same padding, and in this case, padding = 1 and stride = 1, although the size of the output feature is the same as that of the input feature, but it can better preserve the edge information of the input feature.

[0055] Based on this, the convolution module comprises, in sequence, an initial convolution layer, a normalization BN layer, an activation function Relu layer, a maximum pooling layer, a first convolution layer, a second convolution layer, a third convolution layer, a fourth convolution layer and a fifth convolution layer.

[0056] The pooling process is similar to the convolution layer, and also has parameters such as a pooling size, a step and padding. Commonly used pooling methods include mean pooling and maximum pooling, which are to take a mean value and a maximum value in a pooling region respectively. The application adopts maximum pooling to reduce the dimension of the one-dimensional distance image feature after the initial convolution layer. Pooling can effectively reduce the spatial size of data, so the number of parameters and the amount of calculation are also reduced, which also controls overfitting to a certain extent. The result after the pooling operation is reduced compared with the input. The introduction of the pooling layer is to imitate the human visual system to reduce the dimension and abstract the input object. The specific effects of the pooling layer are as follows:

[0057] A) feature invariance: the pooling operation makes the model pay more attention to whether there is a certain feature rather than the specific position of the feature. B) feature dimension reduction (downsampling): pooling is equivalent to dimension reduction in the spatial range, so that the model can extract more wide-range features. At the same time, the input size of the next layer is reduced, thereby reducing the amount of calculation and the number of parameters. C) prevent overfitting to a certain extent, which is more convenient for optimization. D) realize nonlinearity (similar to the activation function relu). E) expand the receptive field.

[0058] After the initial convolution layer, a maximum pooling layer is passed, and then five convolution layers with 64 input channels, 64 output channels, a convolution kernel size of 3, padding = 1 and a step of 1 are passed to complete the feature extraction of the convolution module. The parameter details of each stage are shown in Table 1.

[0059] Table 1 Parameter details of each stage of the convolution module

[0060]

[0061]

[0062] In Table 1, the input channel represents the number of input feature channels. The output channel represents the number of output channels, and the number of output channels requires the number of convolution kernels. The kernel size represents the size of the kernel; if it is a convolution kernel, the size is (k), and the second dimension is determined by the input channel. The padding represents whether to fill, and the default is 0, which means not to fill.

[0063] In summary, through the convolution module, the network learns the features equivalent to the final filter number, i.e., the channel number, so that the training of the network model is more abundant. The output features of this sample are visualized and a heat map is drawn. In the heat map, the horizontal axis represents the feature dimension, and the vertical axis represents the feature channel dimension. As can be seen from the heat map, the output features after the convolution filter pay more attention to the local characteristics of the target, and the support area is more concentrated and significant.

[0064] (ii) Position encoding module

[0065] In order to fully utilize the time sequence relationship of the one-dimensional range image feature sequence, a general cyclic mechanism is usually used to capture position information. RNN, LSTM and other recurrent neural networks capture relative position information along the time dimension through the structure of the input sequence. However, this cyclic structure also makes it difficult for the neural network to perform parallel computation. The present application adds time sequence information to the one-dimensional range image by using position encoding, avoiding the problem of difficult parallel computation of the cyclic structure. From the perspective of signal processing, position encoding is equivalent to adding a positioning signal. Ideally, the design of position embedding should satisfy the following conditions:

[0066] 1) should output a unique code for each unit. 2) the difference between any two units should remain consistent between vectors of different lengths. 3) the value should be bounded.

[0067] Specifically, the input of the model is enhanced by injecting position information for each unit. From a code perspective, it is the sum of position embedding and unit embedding. In this application, the position encoding is encoded using trigonometric functions, and the sin and cos functions are used to encode the odd and even positions of the feature sequence, respectively.

[0068] The position encoding calculation formula is:

[0069]

[0070] where PE(pos, l) represents the lth element in the posth feature dimension, l [0, d model ), d model represents the length of the position encoding, which is equal to L (the number of position embeddings). The value in each feature dimension corresponds to the parity of the embedded position sequence number, for example, 0 and 1 are a group, 2 and 3 are a group, and the sin and cos functions in the formula are used to process them, respectively, to produce different periodic changes. The position embedding will slow down the period as l increases, and eventually produce a texture containing position information.

[0071] According to the properties of the sine and cosine functions, the position encoding vector PE(pos+i) of the (pos+i)-th feature dimension can be expressed as a linear combination of the position encoding vector PE(pos) of the (pos)-th feature dimension and the position encoding vector PE(i) of the i-th feature dimension:

[0072]

[0073] The sequence feature of the position-encoded radar HRRP data is obtained by adding the sequence feature output by the convolution module and the position encoding, and the calculation formula is as follows:

[0074] F pos_emb (pos, l) = F v (pos, l) + PE(pos, l)

[0075] wherein F pos_emb (pos, l) represents the second feature map sequence obtained after position encoding, and F v (pos, l) represents the first feature map sequence.

[0076] The period of the position embedding function varies from 2π to 10000*2π, and each position obtains a combination of values of sin and cos functions with different periods, thereby generating unique texture position information, and the relative position relationship between the sequence features of each channel can be deduced, so that the model can utilize the time sequence relationship between the sequence features of each channel, and finally make the model learn the position-dependent relationship and the time sequence characteristics between the HRRP sequences.

[0077] (Three) BCTE module

[0078] The BCTE module comprises a first BCTE submodule, a second BCTE submodule, a third BCTE submodule and a fourth BCTE submodule connected in sequence; the first BCTE submodule, the second BCTE submodule, the third BCTE submodule and the fourth BCTE submodule have the same module structure, and the first BCTE submodule comprises a multi-head attention mechanism unit, a first residual and normalization unit, a Bayesian linear unit and a second residual and normalization unit.

[0079] The multi-head attention mechanism unit is used to capture the relationship between different feature maps in the second feature map sequence to obtain an integrated sub-feature map; the input end of the multi-head attention mechanism unit is connected with the output end of the position encoding module, and the output end of the multi-head attention mechanism unit is connected with the first input end of the first residual and normalization unit; the output end of the position encoding module is connected with the second input end of the first residual and normalization unit in a skip connection manner.

[0080] The output end of the first residual and normalization unit is connected with the input end of the Bayesian linear unit, and the output end of the Bayesian linear unit is connected with the first input end of the second residual and normalization unit; the output end of the first residual and normalization unit is jump-connected with the second input end of the second residual and normalization unit; and the output end of the second residual and normalization unit is connected with the input end of the second BCTE sub-module.

[0081] Specifically, as follows: The transformer is originally designed to solve the machine translation problem, which is a sequence-to-sequence task. The transformer follows the architecture of the SeqToSeq model: Encoder -> Feature Vector -> Decoder. But unlike the RNN, which is also a cyclic algorithm using the SeqToSeq model, there is no cyclic structure in the entire architecture of the transformer. This not only allows the model to be trained in parallel, but also introduces an attention mechanism called self-attention in the encoder, ensuring that each vector in the sequence that passes through the encoder will contain the context information of the sequence. This makes the feature vector output by the encoder contain much more information than the context information that can be expressed by the feature vector in the cyclic algorithm. Given these advantages, the application selects the structure of the transformer to realize one-dimensional range image sequence classification.

[0082] The linear layer is a fully connected layer, which can be used as a classifier in CNN. CNN can map the original HRRP to the hidden feature space, and the linear layer maps the learned "distributed feature representation" to the sample label space, which uses linear regression. In the frequency-based view, the weight w is an unknown constant, so the problem is converted into an optimization problem, and the weight is point-estimated. The method of point estimation is divided into maximum likelihood estimation and maximum a posteriori estimation. In the Bayesian view, the weight w is a random variable, so the conditional distribution of the random variable is solved. Bayesian linear regression is a linear regression model solved by Bayesian probability inference method, which has the basic properties of Bayesian statistical model. The Bayesian learning process is to multiply the likelihood estimate of the new test sample point by the posterior probability of the previous training set, to obtain the posterior probability of the new set. The posterior probability of the previous set becomes the prior probability distribution of the posterior probability of the new set.

[0083] In the BCTE layer, the weight follows a distribution p(w|T, Y), which satisfies the following Bayesian formula:

[0084]

[0085] where w is the set of model parameters, T is the one-dimensional range image dataset, and Y is the target class label. Because p(w|T, Y) is the probability of w given the data (T, Y), it is difficult to calculate using the above Bayesian formula. Therefore, the present application uses a variational inference method to approximate the posterior probability p(w|T, Y). Specifically, the KL divergence is used to measure the similarity between the variational distribution and the posterior probability p(w|T, Y), and through mathematical derivation, it can be obtained that the evidence lower bound (ELBO) needs to be maximized.

[0086] T i represents the ith one-dimensional range image in the entire one-dimensional range image dataset T. After the model training is completed, each input one-dimensional range image has been predicted H times. Each time a polynomial conditional probability distribution (CPD) p(Y i |T i ,w t ) = MN t∈[1,2,3,4,...,H]. MN under each CNN-BTEN corresponds to a sampling weight w t For each class m∈[1,2,3,...,c], the average probability is calculated according to the following formula:

[0087]

[0088] Then, the highest average probability max(p m (T i ,w)) is used to predict the class of the target, and is calculated according to the following formula:

[0089]

[0090] By Figure 6 It is intuitively shown that any uncertainty is expressed in terms of inter-class distribution, and if the probability of a class is 1, then the uncertainty is 0. Any uncertainty is expressed in terms of the distribution of a class prediction probability, and if the distribution is zero, then the probability is zero. Therefore, CNN-BTEN can obtain more reliable prediction probabilities through the calculation formula of p pred , which captures any and cognitive uncertainty. This advantage can be further demonstrated in the experimental section.

[0091] Compared with linear regression, Bayesian linear regression not only can solve the problem of overfitting in maximum likelihood estimation, but also has a utilization rate of 100% of the data sample, and can effectively and accurately determine the complexity of the model only by using the training sample, but its training overhead is too large. The recognition of one-dimensional distance image is a classification task, the input is a one-dimensional sequence, and the output is the target type which can be represented by a number. Unlike the machine translation problem, it does not need to generate and output a new sequence. In view of the above advantages of Bayesian linear regression, the application uses BayesianLinear Layer and SoftMax as the decoder. Because the attention mechanism can help the model to give different weights to each part of the input sequence feature, extract more key and important information, and the weighted sequence feature can make the model make more accurate judgments, and at the same time will not bring greater overhead to the calculation and storage of the model, which is the reason why the application additionally adds an attention sub-module to the decoder.

[0092] The Encoder in the Transformer encoder mainly consists of Multi-Head Attention and FeedForward layers. Multi-Head Attention is used to capture the relationship between different data points in the input sequence, and expand the ability of the model to focus on different positions. The essence is that multiple independent Attention calculations are performed, and each Attention is responsible for only one subspace in the final output sequence. As an ensemble, it can effectively prevent overfitting. FeedForward consists of two Fully Connected (FC) layers, which first increase the dimensions of the data and then reduce them to be consistent with the input sequence.

[0093] Further, the input feature dimension of the first BCTE sub-module is the same as the output feature dimension; the input feature dimension of the second BCTE sub-module is the same as the output feature dimension of the first BCTE sub-module; the output feature dimension of the second BCTE sub-module, the input and output feature dimensions of the third BCTE sub-module, and the input feature dimension of the fourth BCTE sub-module are all d1; and the output feature dimension of the fourth BCTE sub-module is d2, and d2 is greater than d1.

[0094] Through the above setting, more abundant information can be provided for the calculation of attention, and the encoder layer can better extract feature information between different points in the input sequence. After the input sequence passes through a series of BCTE sub-modules and then passes through the classifier, a completed feature vector for discrimination is obtained. The feature vector contains the probabilities of the one-dimensional distance image belonging to a certain type of ship. The dimensions of d1 and d2 have a very important influence on the recognition ability of the model. If the dimension is too high, it may produce redundancy and increase the training time. If the dimension is too low, important feature information may be lost. The optimal values of d1 and d2 are selected through experiments, d1 is 10, and d2 is 30.

[0095] (iv) Classification module

[0096] The classification module comprises an attention sub-module, a Bayesian linear sub-module and a normalization sub-module connected in sequence; the attention sub-module is used for feature extraction on the third feature map based on an attention mechanism to obtain an attention sub-feature map; and the Bayesian linear sub-module and the normalization sub-module are used as a decoder to perform decoding processing on the attention sub-feature map to obtain a ship type.

[0097] Step 300: training the CNN-BTEN network by using the training sample set to obtain an optimal CNN-BTEN network; and the optimal CNN-BTEN network is a target ship identification model.

[0098] A loss function is an effective tool for measuring the prediction performance of a model, and is used to evaluate the degree of difference between the predicted value and the true value of the model. The better the loss function, the better the performance of the model. Different models generally use different loss functions. The role of the loss function is to calculate the difference between the forward calculation result of each iteration of the neural network and the true value, so as to guide the next training in which direction. The loss function used in the present application is a cross-entropy loss function. CrossEntropyLoss can be directly called in Pytorch, and its calculation formula is as follows:

[0099]

[0100] where log is base e. The first parameter x is the input, which is also the output of the last layer of the network, and its shape is [batchsize, class]. The function requires the first parameter, which is the output of the last layer, to be two-dimensional data, and the value in each vector is the probability value of different categories. The second parameter is the label passed in, that is, the index value of a certain category, which does not participate in the calculation in the formula. If the batchsize is 1, there is only one number, 0 or 1 or 2, etc., indicating the real category corresponding to the sample at this time; if the batchsize is 2, there are two numbers, for example, [0, 1], which respectively represent the real categories corresponding to the two samples. In Pytorch, to get the final cross-entropy loss, it also needs to go through the loss weight, and its calculation formula is as follows:

[0101]

[0102] Secondly, in terms of optimizer selection, Adam optimizer is selected, which combines the advantages of Monument, AdaGrad and RMSProp three optimization algorithms, and through the calculation of the first moment estimate and the second moment estimate of the gradient, it designs independent adaptive learning rate for different parameters, which has the advantages of simple implementation, efficient calculation, less memory demand, suitable for unstable objective function, etc.

[0103] Finally, in terms of learning rate adjustment, OneCycleLR method is selected, that is, the learning rate is first increased in the initial training process to make the model learn faster, and the learning rate is slowly reduced when the training reaches a certain stage to make the model find the global optimal solution.

[0104] Step 400, obtaining a radar high-resolution range profile to be identified.

[0105] Step 500, inputting the radar high-resolution range profile to be identified into the target ship identification model to obtain a ship type in the radar high-resolution range profile to be identified.

[0106] In another specific practical application, the experimental environment and development platform used are shown in Table 2, Pytorch is used for modeling, CPU is used to read data, RTX 3090 is used to train the model, and enough large video memory is used to support simultaneous training of multiple samples to speed up the training speed.

[0107] Table 2 Experimental environment

[0108]

[0109] (1) Similar model comparison under different signal-to-noise ratios.

[0110] CNN-BERT is a neural network mainly using BERT to identify HRRP. First, CNN-BERT passes through a convolution layer with a convolution kernel dimension of (1, 5), a convolution channel number of 768, and a step length of 1, and then passes through a normalization BN layer and a Relu activation function layer; then, positional encoding is performed by using Positional Encoding; then, feature expression is learned by using a stacked 12-layer BERT; finally, a classifier composed of attention and Softmax outputs the probability of belonging to each target category.

[0111] CNN-BTEN is similar to CNN-BERT, and mainly uses BCTE to learn HRRP features. CNN-BTEN passes through a convolution layer with a convolution kernel dimension of (1, 3), a convolution channel number of 64, a step length of 1, and padding set to 1, and then passes through a normalization BN layer and a Relu activation function layer, and then passes through a maximum pooling layer with a pooling kernel dimension of (1, 2) and a step length of 1, and then passes through a convolution module composed of 5 convolution layers with a convolution kernel dimension of (64, 2), a convolution channel number of 64, a step length of 1, and padding set to 1; then, positional encoding is performed by using Positional Encoding; then, feature expression is learned by using a stacked 4-layer BCTE; finally, a classifier composed of attention, BayesianLinear Layer and Softmax outputs the probability of belonging to each target category.

[0112] ModelX is an intermediate comparison item of CNN-BERT and CNN-BTEN. ModelX passes through a convolution layer with a convolution kernel dimension of (1, 5), a convolution channel number of 768, and a step length of 1, and then passes through a normalization BN layer and a Relu activation function layer; then, positional encoding is performed by using Positional Encoding; then, feature expression is learned by using a stacked 4-layer BCTE which is the same as CNN-BTEN; finally, a classifier composed of attention, BayesianLinear Layer and Softmax outputs the probability of belonging to each target category.

[0113] The application selects CNN-BERT, CNN-BTEN and ModelX to perform similar model comparison experiments. The target recognition rates of the three models under different signal-to-noise ratios are shown in Table 3

[0114] Table 3 Target recognition rates (%) of similar models under different signal-to-noise ratios

[0115]

[0116] From Table 3, it can be seen that the target recognition rates of CNN-BERT and CNN-BTEN under different signal-to-noise ratios are similar, and the recognition effects are not much different; and ModelX is slightly worse than the two, and it can be proved that the effect of the convolution module of CNN-BTEN in feature extraction is better than that of a convolution layer with a convolution kernel dimension (1, 5), a channel number 798 and a step length 1.

[0117] From Figure 7 It can be seen that, because CNN-BERT stacks 12 layers of BERT for feature learning, a certain amount of time is consumed. Under the same number of training data and test data, batch size is 64, initial learning rate, learning rate, and the same optimizer, the training time and test time of CNN-BERT are greater than those of CNN-BTEN and ModelX. Although CNN-BTEN has several more convolution layers, the channel number is reduced, so the training time and test time of ModelX are slightly greater than those of CNN-BTEN.

[0118] In summary, both CNN-BERT and CNN-BTEN have high recognition accuracy, and the two are not much different in recognition accuracy, but the training overhead and validation test overhead of CNN-BERT are slightly higher than those of CNN-BTEN. Although ModelX takes less time than CNN-BERT, its recognition accuracy is not as good as that of CNN-BERT. Therefore, it can be seen that the recognition effect of CNN-BTEN is remarkable, the recognition overhead is reasonable, and the model is effective.

[0119] (2) Comparison of target recognition effects of different models under different signal-to-noise ratios.

[0120] The present application mainly compares the recognition effects of the four models of CNN-BTEN, RNN, GMM and SVM. From Table 4, it can be seen that the overall recognition effect of the six types of ship data under different signal-to-noise ratios is better than that of the other models.

[0121] Table 4 Target recognition rate (%) of different models under different signal-to-noise ratios

[0122]

[0123]

[0124] From Table 5, it can be seen that the recognition effect of the six types of ship data under 10dB signal-to-noise ratio, the recognition effect of CNN-BTEN for single class is balanced, and on average it is better than the other models. It can be seen that the CNN-BTEN model has better target recognition performance and better robustness, and the accuracy is also guaranteed.

[0125] Table 5 Recognition rate of different models for single-class targets under 10dB signal-to-noise ratio

[0126]

[0127] Meanwhile, the ship data added with 10dB noise is plotted to obtain the confusion matrix of each model. By plotting the confusion matrix of each model, the detailed conditions of correct classification and error classification of the six types of sea surface targets can be seen, and it can also be known that 12.91% of the ships are classified into eta by SVM, and the error rate is relatively high. The confusion matrix also reflects from the side whether the distribution of the six types of ship data learned by the model is similar, and can effectively compare the advantages and disadvantages of the classifier.

[0128] (3) Feature scatter plot analysis

[0129] Commonly used dimension reduction visualization methods include the dimension reduction PCA (Principal Component Analysis) algorithm and the LDA (Linear Discriminant Analysis) algorithm based on supervised learning for dimension reduction. However, these two algorithms have certain defects, and the experimental effect of the data of the present application cannot reflect the distinct classification effect of the features output by each module in two-dimensional space. Therefore, the t-SNE algorithm is selected for dimension reduction to view the classification effect. The t-SNE (t-Distributed Stochastic Neighbor Embedding) is a machine learning algorithm for reducing dimension, which is suitable for visualizing high-dimensional data. The main goal of t-SNE is to convert a multidimensional data set into a low-dimensional data set. Compared with other dimension reduction algorithms, t-SNE has the best effect for data visualization. If t-SNE is applied to n-dimensional data, it will intelligently map the n-dimensional data to 3d or even 2d data, and the relative similarity of the original data is very good. Like PCA, t-SNE is not a linear dimension reduction technique, and it follows a nonlinear, which is the main reason why it can capture the complex manifold structure of high-dimensional data.

[0130] The present application selects the t-SNE algorithm to perform scatter plot distribution visualization on the features output by each module of the CNN-BTEN. Through three-point distribution visualization, it can be known that the originally disordered data gradually begins to converge after being filtered by the convolution module; then the classification has a preliminary significant effect after the repeated scaling and learning features of the four BCTE sub-modules; due to the large amount of prior information obtained by training, combined with the current obtained data, the information about unknown parameters is updated, so as to obtain the possibility of revised parameters, that is, the posterior probability. Finally, through the classifier, the features are mapped to the hidden layer feature space through the Bayesian Linear layer.

[0131] In summary, the application proposes a new deep learning framework CNN-BTEN for radar HRRP data recognition, which uses multiple convolutional layers and pooling layers to filter one-dimensional range images to obtain features that focus more on local features. Secondly, the position of the feature is encoded. In order to fully utilize the time sequence relationship of the one-dimensional range image feature sequence, the general method usually uses a cyclic mechanism to capture position information. RNN, LSTM and other recurrent neural networks capture relative position information along the time dimension through the structure of the input sequence, however, this cyclic structure also makes it difficult for recurrent neural networks to parallel computing. The application uses position encoding to add timing information to HRRP, avoiding the problem of difficult parallel computing of the cyclic structure. At the same time, the BCTE module is used to extract features to obtain prior and posterior information of HRRP and information containing various probability distributions. Finally, after the attention mechanism, Bayesian linear layer and normalization exponential function of the classifier module, the target recognition probability of each target is obtained, and the final target classification is completed. In addition, in the training process of the network, the OneCycleLR version of CyclicLR is used to adjust the learning rate, so that the model can converge more quickly and effectively.

[0132] The experiment is verified by using 6 types of simulation data under different noise conditions, and compared with various existing models, under 10dB Gaussian noise, CNN-BTEN can reach 99.08% recognition rate compared with RNN, AE and SVM. The experimental results show the superior performance of the CNN-BTEN model. When designing the learning rate, the OneCycleLR method is used, which first increases the learning rate in the initial training process to make the model learn faster, and slowly reduces the learning rate when training to a certain stage to make the model find the global optimal solution, so as to realize stable and efficient.

[0133] Embodiment two

[0134] As Figure 8 shown, to realize the scheme as recorded in embodiment one, the embodiment provides an intelligent radar HRRP target recognition system, which comprises:

[0135] A training sample determination subsystem 101 is configured to obtain a training sample set, wherein each training sample in the training sample set comprises a radar high-resolution range sample image and a ship type corresponding to the radar high-resolution range sample image.

[0136] The network construction subsystem 201 is configured to construct a CNN-BTEN network; the CNN-BTEN network comprises a convolution module, a position coding module, a BCTE module and a classification module; the convolution module is configured to perform feature extraction on multi-dimensional range images in the radar high-resolution range sample image respectively to obtain a first feature map sequence; the position coding module is configured to perform position coding on the first feature map sequence by using a trigonometric function to obtain a second feature map sequence; the BCTE module is configured to perform feature extraction on the second feature map sequence based on a Bayesian linear regression and a Transformer encoder to obtain a third feature map; and the classification module is configured to calculate the probability of each ship type in the third feature map to realize ship target classification.

[0137] The network training subsystem 301 is configured to train the CNN-BTEN network by using the training sample set to obtain an optimal CNN-BTEN network; and the optimal CNN-BTEN network is a target ship identification model.

[0138] The to-be-identified image acquisition subsystem 401 is configured to acquire a to-be-identified radar high-resolution range image.

[0139] The image identification subsystem 501 is configured to input the to-be-identified radar high-resolution range image into the target ship identification model to obtain a ship type in the to-be-identified radar high-resolution range image.

[0140] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments, and the same or similar parts between the embodiments can be referred to each other.

[0141] The principles and implementation manners of the present application are described by using specific examples in the present application, and the above description of the embodiments is only used to help understand the method of the present application and its core idea; meanwhile, for those skilled in the art, according to the idea of the present application, the specific implementation manner and application range can be changed. In conclusion, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A smart radar HRRP target recognition method, characterized in that, The method comprises: obtaining a training sample set; each training sample in the training sample set comprises a radar high-resolution range sample image and a corresponding ship type of the radar high-resolution range sample image; constructing a CNN-BTEN network; the CNN-BTEN network comprises a convolution module, a position encoding module, a BCTE module and a classification module; the convolution module is used for performing feature extraction on multi-dimensional range images in the radar high-resolution range sample image respectively to obtain a first feature map sequence; the position encoding module is used for performing position encoding on the first feature map sequence by using a trigonometric function to obtain a second feature map sequence; the BCTE module is used for performing feature extraction on the second feature map sequence based on Bayesian linear regression and a Transformer encoder to obtain a third feature map; and the classification module is used for calculating the probability of each ship type in the third feature map to realize ship target classification; a calculation formula of position encoding in the position encoding module is as follows: wherein, represents a second feature map sequence obtained after position encoding, represents a first feature map sequence, represents a first pos position encoding vector in a first feature dimension, represents a first position encoding vector in a first feature dimension, ), represents the length of position encoding; the BCTE module comprises a first BCTE submodule, a second BCTE submodule, a third BCTE submodule and a fourth BCTE submodule connected in sequence; the first BCTE submodule, the second BCTE submodule, the third BCTE submodule and the fourth BCTE submodule have the same module structure, and the first BCTE submodule comprises a multi-head attention mechanism unit, a first residual and normalization unit, a Bayesian linear unit and a second residual and normalization unit; the multi-head attention mechanism unit is used for capturing the relationship between different feature maps in the second feature map sequence to obtain an integrated sub-feature map; an input end of the multi-head attention mechanism unit is connected with an output end of the position encoding module, an output end of the multi-head attention mechanism unit is connected with a first input end of the first residual and normalization unit, the output end of the position encoding module is jump-connected with a second input end of the first residual and normalization unit, an output end of the first residual and normalization unit is connected with an input end of the Bayesian linear unit, an output end of the Bayesian linear unit is connected with a first input end of the second residual and normalization unit, the output end of the first residual and normalization unit is jump-connected with a second input end of the second residual and normalization unit, and an output end of the second residual and normalization unit is connected with an input end of the second BCTE submodule; training the CNN-BTEN network by using the training sample set to obtain an optimal CNN-BTEN network; the optimal CNN-BTEN network is a target ship identification model; obtaining a radar high-resolution range image to be identified; inputting the radar high-resolution range image to be identified into the target ship identification model to obtain a ship type in the radar high-resolution range image to be identified. 2.The intelligent radar HRRP target recognition method of claim 1, wherein, The convolution module comprises an initial convolution layer, a normalization BN layer, an activation function Relu layer, a maximum pooling layer, a first convolution layer, a second convolution layer, a third convolution layer, a fourth convolution layer and a fifth convolution layer connected in sequence. 3.The intelligent radar HRRP target recognition method of claim 1, wherein, The input feature dimension of the first BCTE submodule is the same as the output feature dimension; the input feature dimension of the second BCTE submodule is the same as the output feature dimension of the first BCTE submodule; The output feature dimension of the second BCTE submodule, the input and output feature dimensions of the third BCTE submodule, and the input feature dimension of the fourth BCTE submodule are all d1; The output feature dimension of the fourth BCTE submodule is d2, and d2 is greater than d1.

4. The intelligent radar HRRP target recognition method according to claim 3, characterized in that, The Bayesian linear unit inner weights follow distribution; the distribution is determined by variational inference; w is a set of model parameters, T is a one-dimensional distance image dataset, and Y is a target class label.

5. The intelligent radar HRRP target recognition method according to claim 1, characterized in that, The classification module comprises an attention submodule, a Bayesian linear submodule and a normalization submodule connected in sequence; The attention submodule is configured to perform feature extraction on the third feature map based on an attention mechanism to obtain an attention sub-feature map; The Bayesian linear submodule and the normalization submodule are configured to decode the attention sub-feature map to obtain a ship type. 6.The intelligent radar HRRP target recognition method of claim 1, wherein, The training sample set is obtained by: constructing ship models of multiple types; using FEKO software to simulate the ship models of the multiple types to obtain corresponding multiple initial radar high-resolution range images; performing data enhancement processing or multiple times of adding Gaussian white noise processing on each of the initial radar high-resolution range images to obtain multiple radar high-resolution range sample images; determining the ship type of the ship model in each radar high-resolution range sample image; each radar high-resolution range sample image and the ship type corresponding to the radar high-resolution range sample image constitute a single training sample; and multiple training samples constitute a training sample set.

7. The intelligent radar HRRP target recognition method according to claim 1, characterized in that, The method further comprises: adjusting the learning rate of the CNN-BTEN network using the OneCycleLR method during the training of the CNN-BTEN network using the training sample set.

8. An intelligent radar HRRP target recognition system characterized in that, The system comprises: a training sample determination subsystem configured to obtain a training sample set; each training sample in the training sample set comprises a radar high-resolution range sample image and a ship type corresponding to the radar high-resolution range sample image; a network construction subsystem configured to construct a CNN-BTEN network; the CNN-BTEN network comprises a convolution module, a position encoding module, a BCTE module and a classification module; the convolution module is configured to perform feature extraction on multiple-dimensional range images in the radar high-resolution range sample image to obtain a first feature map sequence; the position encoding module is configured to perform position encoding on the first feature map sequence using a trigonometric function to obtain a second feature map sequence; the BCTE module is configured to perform feature extraction on the second feature map sequence based on Bayesian linear regression and a Transformer encoder to obtain a third feature map; and the classification module is configured to calculate the probability of each ship type in the third feature map to realize ship target classification; The calculation formula of the position encoding in the position encoding module is as follows: wherein, denotes the second feature map sequence obtained after position encoding, denotes the first feature map sequence, denotes the pos th position encoding vector in the th feature dimension, ∈ [0, ), denotes the length of the position encoding; The BCTE module comprises a first BCTE submodule, a second BCTE submodule, a third BCTE submodule and a fourth BCTE submodule connected in sequence; the first BCTE submodule, the second BCTE submodule, the third BCTE submodule and the fourth BCTE submodule have the same module structure, and the first BCTE submodule comprises a multi-head attention mechanism unit, a first residual and normalization unit, a Bayesian linear unit and a second residual and normalization unit; The multi-head attention mechanism unit is used for capturing the relationship between different feature maps in the second feature map sequence to obtain integrated sub-feature maps; an input end of the multi-head attention mechanism unit is connected with an output end of the position encoding module, an output end of the multi-head attention mechanism unit is connected with a first input end of the first residual and normalization unit, and the output end of the position encoding module is jump-connected with a second input end of the first residual and normalization unit; an output end of the first residual and normalization unit is connected with an input end of the Bayesian linear unit, an output end of the Bayesian linear unit is connected with a first input end of the second residual and normalization unit, and an output end of the first residual and normalization unit is jump-connected with a second input end of the second residual and normalization unit; an output end of the second residual and normalization unit is connected with an input end of the second BCTE submodule; A network training subsystem is configured to train the CNN-BTEN network by using the training sample set to obtain an optimal CNN-BTEN network; the optimal CNN-BTEN network is a target ship identification model; A to-be-identified image acquisition subsystem is configured to acquire a to-be-identified radar high-resolution range profile; An image identification subsystem is configured to input the to-be-identified radar high-resolution range profile into the target ship identification model to obtain a ship type in the to-be-identified radar high-resolution range profile.

Images