A Class- and Feature-Assisted Neural Network-Based Ultra-Narrow Pulse Radar Target Recognition Method

Abstract

The present invention relates to a method for identifying extremely narrow pulse radar targets using a category and feature-assisted neural network, and belongs to the field of radar target recognition technology. In particular, it relates to a method for identifying extremely narrow pulse radar echoes of targets using a deep neural network assisted by prior knowledge. The present invention proposes a method for embedding target electromagnetic feature knowledge into a deep neural network by combining a bag-of-words model and an attention mechanism. This method integrates and represents target feature knowledge through a bag-of-words model, making the representation of feature knowledge and deep features more similar, which is conducive to the combination of the two types of information. At the same time, the attention mechanism is used to enable the network to adaptively focus on feature combinations that contribute more to the classification task, thereby enhancing the discriminability of feature combinations and further improving the recognition performance of the network.

Description

Technical Field

[0001] The present invention relates to a method for recognizing extremely narrow pulse radar targets using a category and feature-assisted neural network, belonging to the technical field of radar target recognition, and in particular to a method for recognizing extremely narrow pulse radar echoes of targets using a deep neural network assisted by prior knowledge. Background Art

[0002] Ultra-narrow pulse radars are radars in which, after processing, the pulse width of a single echo is significantly smaller than the target's size. For an ultra-narrow pulse radar, the target's echo consists of multiple extremely narrow pulses, each corresponding to a different scattering center on the target. Therefore, the target's ultra-narrow pulse echo can represent the distribution of the target's scattering centers along the radar's line of sight, often referred to as the target's High Resolution Range Profile (HRRP).

[0003] Traditionally, the method for utilizing target characteristic information involves manually extracting multi-dimensional features of the target, such as high-resolution features, polarization features, and scale features, and then applying these features to classifiers such as support vector machines (SVMs) for recognition. Manual feature extraction relies on extensive practical experience, making it difficult to develop a generalizable solution. Furthermore, the workload of feature analysis and screening is enormous, all of which hinder the recognition performance and application of such methods.

[0004] In recent years, deep learning technology has gradually entered the field of radar target recognition. This approach can autonomously extract discriminative data features, achieve end-to-end target recognition, and achieve good recognition results. However, because this method only fits network parameters based on data and lacks theoretical or logical constraints, it is prone to overfitting, which limits recognition performance. Summary of the Invention

[0005] The technical problem solved by the present invention is: to overcome the shortcomings of the existing technology and propose an ultra-narrow pulse radar target recognition method with a category and feature-assisted neural network. This method can further improve the recognition performance of deep learning methods in the field of radar target recognition. On the one hand, a deep neural network is used to extract information based on target data to obtain target deep features; on the other hand, known useful information or knowledge is used to assist and guide network training, making up for the deficiencies of deep learning in knowledge and logic, guiding and constraining network training, and improving its recognition performance.

[0006] The technical solution of the present invention is:

[0007] A method for ultra-narrow pulse radar target recognition using a category- and feature-assisted neural network is proposed. This method constructs a deep neural network based on the ultra-narrow pulse echoes of radar targets, introduces prior knowledge to assist at different stages of network training, and accurately classifies different targets. The method incorporates the hierarchical category relationships of targets as prior knowledge into the loss function, constraining the updating and optimization of network parameters and improving the network's recognition performance. The method also embeds knowledge of target electromagnetic scattering characteristics into the network, effectively combining it with the deep features extracted by the network to enhance the network's pertinence and discriminability in learning feature representations.

[0008] The specific steps of this method include:

[0009] Step 1: Extract the scattering center of the target extremely narrow pulse echo data and obtain the electromagnetic feature knowledge of the extracted scattering center. Then, use the extracted electromagnetic feature knowledge of the scattering center to generate a target feature representation matrix through the bag-of-words model;

[0010] Step 2: Preprocess the target extremely narrow pulse echo data and use a deep neural network to extract the target depth features from the preprocessed data;

[0011] Step 3: Using the attention mechanism, combine the target feature representation matrix generated in step 1 with the target deep features generated in step 2 to obtain a classification feature vector, and predict the target category based on the classification feature vector;

[0012] Step 4: Construct a knowledge-constrained loss function based on the target category predicted in step 3 and the actual category level prior knowledge of the target to be detected;

[0013] Step 5: Train and update the deep neural network parameters according to the knowledge constraint loss function constructed in step 4 to obtain the trained and updated deep neural network;

[0014] Step 6: Perform target recognition based on the trained and updated deep neural network obtained in step 5 to obtain the target category, completing the ultra-narrow pulse radar target recognition using the category and feature-assisted neural network.

[0015] In the step 1, the scattering centers of the target extremely narrow pulse echo data are extracted using the ordered statistical constant false alarm detection (OS-CFAR) technology, the number of extracted scattering centers is R, and any scattering center is represented by r, where r=1, 2, 3, ..., R;

[0016] There are z kinds of electromagnetic characteristic knowledge of the scattering center r extracted, and 9 kinds are preferably selected, namely position proportion, intensity proportion, sparsity characteristic, scattering polarization entropy H, scattering polarization angle α, power matrix trace, polarization ratio, polarization correlation coefficient, and frequency domain polarization divergence;

[0017] The position ratio indicates the position of the scattering center in the target area, and the length from the target starting point to the scattering center is the percentage LEN of the target area length. r Indicates as follows:

[0018]

[0019] Among them, len r Indicates the position of the scattering center r, len s Indicates the starting distance unit position of the target area, len e Indicates the end distance unit position of the target area;

[0020] The intensity ratio indicates the ratio of the intensity of the scattering center to the total intensity of the target area, which can reflect the energy of the scattering center and its contribution to the target echo, expressed as E r :

[0021]

[0022] Among them, e r represents the scattered power at the scattering center r;

[0023] The sparsity feature indicates the sparsity of the scattering center and the surrounding scattering centers, which can reflect the spatial distribution of the scattering center. The variance of the distance between the rth scattering center and the left and right scattering centers is calculated as the sparsity feature Sp r ,as follows:

[0024] Sp r =var(sp D1 ,sp D2 )

[0025] Among them, sp D1 is the distance between the scattering center r and the left scattering center, sp D2 is the distance between the scattering center r and the right scattering center, var(sp D1 ,sp D2 ) is sp D1 and sp D2 variance;

[0026] The scattering polarization entropy H and the scattering polarization angle α describe the randomness of the scattering of the resolution unit within the extremely narrow pulse echo segment of the target. The scattering polarization angle α describes the dominant radiation mechanism within the extremely narrow pulse echo segment of the target. The three distance units to the left and right of each scattering center are used as segments to calculate the scattering polarization entropy H and the scattering polarization angle α as follows:

[0027]

[0028]

[0029] Where λ1 represents the maximum eigenvalue of the covariance matrix of the target extremely narrow pulse echo segment, λ2 represents the second largest eigenvalue of the covariance matrix of the target extremely narrow pulse echo segment, i = [1, 2], λ i represents the i-th eigenvalue of the covariance matrix of the target extremely narrow pulse echo segment;

[0030] The power matrix trace represents the energy of the scattering center, which is the total power received by a pair of orthogonally polarized antennas:

[0031] Span r =|HH r | 2 +|VV r | 2 +|VH r | 2 +|HV r | 2

[0032] Among them, HH r is the amplitude of the scattering center r in the HH polarization mode;

[0033] VV r is the amplitude of the scattering center r in HV polarization mode;

[0034] VH r is the amplitude of the scattering center r in the VH polarization mode;

[0035] HV r is the amplitude of the scattering center r in HV polarization mode;

[0036] The polarization ratio and polarization correlation coefficient reflect the differences in scattering mechanisms caused by factors such as the structure of the target scattering center:

[0037]

[0038] Polarization correlation coefficient

[0039] Among them, <.> represents spatial average; superscript * represents the conjugate operation; j represents an imaginary number, represents the phase of the scattering center r in the HH polarization mode, represents the phase of the scattering center r in the VH polarization mode, To find the exponential function of the phase of the scattering center r in the HH polarization mode, To find the exponential function of the phase of the scattering center r in the VH polarization mode;

[0040] Frequency domain polarization divergence Div r The density of the spatial distribution of the target echo in the frequency domain transient polarization projection set is described, and the cluster center is obtained from the frequency domain transient polarization projection set of the target echo, and further obtained;

[0041] In the above step 1, for each electromagnetic feature knowledge, a feature representation vector is generated by the word bag model, and all the generated feature representation vectors are combined to form a target feature representation matrix. Since the number of scattering centers contained in each echo data is not fixed and the extracted features are relatively rich, direct use will result in different feature dimensions of different targets and an excessive computational burden. Therefore, the present invention uses the word bag model to form a more stable feature representation vector with unified dimensions for each of the z target features. The specific implementation method is as follows:

[0042] (1) Any electromagnetic feature knowledge is represented by g, where g = 1, 2, ..., z. The bag-of-words model regards different values of the electromagnetic feature knowledge g as different words and obtains k cluster centers through clustering. That is, the electromagnetic feature knowledge g is divided into k different sets according to the differences between the values.

[0043] (2) For each target's extremely narrow pulse echo data, determine to which set the electromagnetic characteristic knowledge g of all scattering centers belongs, that is, to which cluster center it is closer;

[0044] (3) Statistically analyze the judgment results corresponding to each cluster center to obtain the probability value of the electromagnetic characteristic knowledge g of the scattering center of each target extremely narrow pulse echo data belonging to different cluster centers, and use this probability vector with a dimension of 1×k as the feature representation vector of the electromagnetic characteristic knowledge g of the target extremely narrow pulse echo data;

[0045] (4) Repeat (1) to (3) and concatenate the z types of feature representation vectors obtained to obtain a z × k target feature representation matrix for each target extremely narrow pulse echo data;

[0046] In the step 2, the method for preprocessing the target extremely narrow pulse echo data is:

[0047] In order to make full use of the amplitude and phase information of the target extremely narrow pulse echo data, the target extremely narrow pulse echo data X is divided into real and imaginary parts, and then the real and imaginary parts are spliced to form a data vector. Finally, the data vector is normalized to between 0 and 1 and represented by x:

[0048] x=norm([Re(X),Im(X)])

[0049] Wherein, Re(X) represents the operation of taking the real part of the target extremely narrow pulse echo data X, Im(X) represents the operation of taking the imaginary part of the target extremely narrow pulse echo data X, and norm represents the maximum value normalization operation;

[0050] The method for constructing a deep neural network is as follows: the deep neural network includes four network blocks, an average pooling layer, and a linear mapping layer. The first and third network blocks are the same and both contain two convolutional layers. The second and fourth network blocks have the same structure and both contain three convolutional layers. Each network block is constructed using a convolutional layer and a residual connection structure. The residual connection structure is proposed to solve the degradation problem of deep neural networks. It can make networks with more layers converge better and improve the learning ability and stability of the network. For the input x, the output O of the residual connection structure is expressed as:

[0051] O = relu(F(x)+x)

[0052] Among them, relu represents the "relu" nonlinear activation function, and F(x) represents the result obtained by the convolutional layer of the neural network;

[0053] Input the preprocessed data into the constructed deep neural network to obtain the target depth feature f d , f d The dimension is 1×d, where d is a set value and is a known value;

[0054] In step 3, the method of combining the target feature representation matrix with the target depth feature is:

[0055] 301. First, the target feature representation matrix is linearly mapped to z×d to obtain a representation closer to the target depth feature. Then, each row of the reduced target feature representation matrix is concatenated with the target depth feature, that is, each feature representation vector is combined with the target depth feature to obtain a feature combination matrix V with a dimension of z×2d.

[0056] 302. Input the feature combination matrix V obtained in step 301 into the attention mechanism, and redistribute weights to each feature combination vector in the feature combination matrix V. Feature combination vectors with large weights are considered to be able to better distinguish targets, so that the network pays more attention to more discriminative feature combination vectors.

[0057] The method for redistributing weights is as follows: the attention mechanism first uses an average pooling layer and two fully connected layers to learn a weight vector m corresponding to each feature combination vector, and then multiplies the weight vector m with the feature combination matrix V to obtain a weighted feature combination matrix V′;

[0058] The weight vector m is calculated as follows:

[0059] m=Pool(σ(W2δ(W1V)))

[0060] Among them, Pool() represents the average pooling operation, σ represents the "sigmoid" activation function, δ represents the "relu" activation function, and W2 and W1 are network parameters;

[0061] The weighted feature combination matrix V′ is:

[0062] V′=V·m;

[0063] In step 3, the method for predicting the target category based on the classification feature vector is:

[0064] Use two fully connected layers to predict the category of the target. Specifically, first flatten the weighted feature combination matrix V' into a one-dimensional vector, use the dimension P output by the first fully connected layer to reduce the dimension of the weighted feature combination matrix V', and use "relu" as the activation function to obtain the final classification feature vector f cls , the output dimension of the second fully connected layer is the number of target categories, and the activation function is "softmax". The category with the largest output probability is the predicted target category;

[0065] In step 4, the method for constructing the knowledge constraint loss function is:

[0066] 401. A contrast loss function is established based on the classification feature vector obtained in step 3 and the true category of the target to be detected. The contrast loss function is:

[0067]

[0068] Among them, Y represents the difference in the true categories of the two targets to be tested. Y = 1 means that the true categories of the two targets to be tested are the same, and Y = 0 means that the true categories of the two targets to be tested are different. cls1 and f cls2 Represents the classification feature vector of the two targets to be tested, N represents the number of targets to be tested, T is the set threshold, D represents the Euclidean distance between the classification feature vectors of the two targets to be tested, P represents the dimension of the classification feature vector, and the superscript p represents the pth feature in the classification feature vector;

[0069] When Y=1, the contrast loss function only has That is, when the true categories of the two targets to be tested are the same, the Euclidean distance of the classification feature vectors is large, which means that the deep neural network is not well trained;

[0070] When Y=0, the contrast loss function is That is, when the true categories of the two targets to be tested are different, the Euclidean distance of the classification feature vectors is large, which means that the deep neural network is well trained. When the Euclidean distance of the classification feature vectors exceeds T, the contrast loss function is set to 0;

[0071] 402. Using the prior knowledge of the true category hierarchy of the target to be measured, a logical relationship expression is established to improve the contrast loss function constructed in step 401 to obtain an improved contrast loss function;

[0072] The method of establishing a logical relationship expression based on the prior knowledge of the true category level of the target to be tested is:

[0073] Taking u1 and u2 as examples, we first calculate the true category hierarchy number C(u1) of target u1 and the true category hierarchy number C(u2) of target u2. Then we calculate the hierarchy number h of the two targets u1 and u2 whose true categories are the same. The logical relationship expression Y of the two targets is: c Expressed as:

[0074]

[0075] Assume that both targets u1 and u2 belong to the two categories of vehicles and cars, but the models are different. In this case, the number of true category levels is C(u1)=C(u2)=3, and the number of levels of the same true category of targets u1 and u2 is h=2, so we get Y c =2 / 3,

[0076] The improved contrast loss function is:

[0077]

[0078] 403. Combine the improved contrast loss function with the cross entropy loss function to construct the knowledge constraint loss function. The knowledge constraint loss function is:

[0079] L=γ1L c +γ2L CF

[0080] Among them, γ1 and γ2 are proportional coefficients, L C is the improved contrast loss function, L CE is the cross entropy loss function, y n is the true category of the nth target to be tested, is the predicted target category of the nth target to be tested.

[0081] In step 5, the method for training and updating the deep neural network parameters is:

[0082] During the training phase, the input of the deep learning network is the target containing the true category and the corresponding electromagnetic feature knowledge. For any target, the predicted category of the target is obtained through the network forward propagation, and the network loss is calculated using the knowledge constraint loss function L. Then, the network parameters are updated layer by layer through the Adam optimizer and the backpropagation algorithm to minimize the loss function and make the network converge, thus obtaining the trained and updated deep neural network.

[0083] In step 6, the method for performing target identification is:

[0084] The input of the deep learning network is the target of unknown true category and the corresponding electromagnetic feature knowledge. The target category is predicted through the trained and updated deep neural network. The category with the highest probability output by the last layer is the target category.

[0085] Compared with the existing technology, the present invention has the following advantages:

[0086] (1) This paper proposes a knowledge-constrained loss function based on prior knowledge of the target hierarchical categories. This loss function is constructed based on label information and network-learned features, and further introduces known hierarchical knowledge of the target categories to further constrain the training process of the network parameters, narrowing the solution space of the neural network parameters and improving the network recognition accuracy.

[0087] (2) The present invention proposes a method for embedding target electromagnetic feature knowledge into a deep neural network by combining a bag-of-words model and an attention mechanism. This method integrates and represents target feature knowledge through a bag-of-words model, making the representation of feature knowledge more similar to that of deep features, which is conducive to the combination of the two types of information. At the same time, the attention mechanism is used to enable the network to adaptively focus on feature combinations that contribute more to the classification task, thereby enhancing the discriminability of feature combinations and further improving the recognition performance of the network.

[0088] (3) Compared with existing methods, the method of the present invention uses the bag-of-words model to narrow the gap between feature knowledge and deep features, and uses the attention mechanism to enhance the discriminability of feature combinations. It can better embed feature knowledge into the neural network and make full use of feature knowledge and deep feature information.

[0089] (4) The proposal and implementation of the present invention has high promotion and application value in the field of radar target extremely narrow pulse echo recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0090] Figure 1 The true category-level prior knowledge of 10 types of targets to be detected;

[0091] Figure 2 This is the confusion matrix diagram of the classification results. DETAILED DESCRIPTION

[0092] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0093] A method for ultra-narrow pulse radar target recognition using a category and feature-assisted neural network, comprising the following steps:

[0094] Step 1: Extract the scattering center of the target extremely narrow pulse echo data and obtain the electromagnetic feature knowledge of the extracted scattering center. Then, use the extracted electromagnetic feature knowledge of the scattering center to generate a target feature representation matrix through the bag-of-words model;

[0095] Step 2: Preprocess the target extremely narrow pulse echo data and use a deep neural network to extract the target depth features from the preprocessed data;

[0096] Step 3: Using the attention mechanism, combine the target feature representation matrix generated in step 1 with the target deep features generated in step 2 to obtain a classification feature vector, and predict the target category based on the classification feature vector;

[0097] Step 4: Construct a knowledge-constrained loss function based on the target category predicted in step 3 and the actual category level prior knowledge of the target to be detected;

[0098] Step 5: Train and update the deep neural network parameters according to the knowledge constraint loss function constructed in step 4 to obtain the trained and updated deep neural network;

[0099] Step 6: Perform target recognition based on the trained and updated deep neural network obtained in step 5 to obtain the target category, completing the ultra-narrow pulse radar target recognition using the category and feature-assisted neural network.

[0100] In the step 1, the scattering centers of the target extremely narrow pulse echo data are extracted using the ordered statistical constant false alarm detection (OS-CFAR) technology, the number of extracted scattering centers is R, and any scattering center is represented by r, where r=1, 2, 3, ..., R;

[0101] There are z kinds of electromagnetic characteristic knowledge of the scattering center r extracted, and 9 kinds are preferably selected, namely position proportion, intensity proportion, sparsity characteristic, scattering polarization entropy H, scattering polarization angle α, power matrix trace, polarization ratio, polarization correlation coefficient, and frequency domain polarization divergence;

[0102] The position ratio indicates the position of the scattering center in the target area, and the length from the target starting point to the scattering center is the percentage LEN of the target area length. r Indicates as follows:

[0103]

[0104] Among them, len rIndicates the position of the scattering center r, len s Indicates the starting distance unit position of the target area, len e Indicates the end distance unit position of the target area;

[0105] The intensity ratio indicates the ratio of the intensity of the scattering center to the total intensity of the target area, which can reflect the energy of the scattering center and its contribution to the target echo, expressed as E r :

[0106]

[0107] Among them, e r represents the scattered power at the scattering center r;

[0108] The sparsity feature indicates the sparsity of the scattering center and the surrounding scattering centers, which can reflect the spatial distribution of the scattering center. The variance of the distance between the rth scattering center and the left and right scattering centers is calculated as the sparsity feature Sp r ,as follows:

[0109] Sp r =var(sp D1 ,sp D2 )

[0110] Among them, sp D1 is the distance between the scattering center r and the left scattering center, sp D2 is the distance between the scattering center r and the right scattering center, var(sp D1 ,sp D2 ) is sp D1 and sp D2 variance;

[0111] The scattering polarization entropy H and the scattering polarization angle α describe the randomness of the scattering of the resolution unit within the extremely narrow pulse echo segment of the target. The scattering polarization angle α describes the dominant radiation mechanism within the extremely narrow pulse echo segment of the target. The three distance units to the left and right of each scattering center are used as segments to calculate the scattering polarization entropy H and the scattering polarization angle α as follows:

[0112]

[0113]

[0114] Where λ1 represents the maximum eigenvalue of the covariance matrix of the target extremely narrow pulse echo segment, λ2 represents the second largest eigenvalue of the covariance matrix of the target extremely narrow pulse echo segment, i = [1, 2], λ i represents the i-th eigenvalue of the covariance matrix of the target extremely narrow pulse echo segment;

[0115] The power matrix trace represents the energy of the scattering center, which is the total power received by a pair of orthogonally polarized antennas:

[0116] Span r =|HH r | 2 +|VV r | 2 +|VH r | 2 +|HV r | 2

[0117] Among them, HH r is the amplitude of the scattering center r in the HH polarization mode;

[0118] VV r is the amplitude of the scattering center r in HV polarization mode;

[0119] VH r is the amplitude of the scattering center r in the VH polarization mode;

[0120] HV r is the amplitude of the scattering center r in HV polarization mode;

[0121] The polarization ratio and polarization correlation coefficient reflect the differences in scattering mechanisms caused by factors such as the structure of the target scattering center:

[0122] Polarization ratio

[0123] Polarization correlation coefficient

[0124] Among them, <.> represents spatial average; superscript * represents the conjugate operation; j represents an imaginary number, represents the phase of the scattering center r in the HH polarization mode, represents the phase of the scattering center r in the VH polarization mode, To find the exponential function of the phase of the scattering center r in the HH polarization mode, To find the exponential function of the phase of the scattering center r in the VH polarization mode;

[0125] Frequency domain polarization divergence Div r The density of the spatial distribution of the target echo in the frequency domain transient polarization projection set is described, and the cluster center is obtained from the frequency domain transient polarization projection set of the target echo, and further obtained;

[0126] In the above step 1, for each electromagnetic feature knowledge, a feature representation vector is generated by the word bag model, and all the generated feature representation vectors are combined to form a target feature representation matrix. Since the number of scattering centers contained in each echo data is not fixed and the extracted features are relatively rich, direct use will result in different feature dimensions of different targets and an excessive computational burden. Therefore, the present invention uses the word bag model to form a more stable feature representation vector with unified dimensions for each of the z target features. The specific implementation method is as follows:

[0127] (1) Any electromagnetic feature knowledge is represented by g, where g = 1, 2, ..., z. The bag-of-words model regards different values of the electromagnetic feature knowledge g as different words and obtains k cluster centers through clustering. That is, the electromagnetic feature knowledge g is divided into k different sets according to the differences between the values.

[0128] (2) For each target's extremely narrow pulse echo data, determine to which set the electromagnetic characteristic knowledge g of all scattering centers belongs, that is, to which cluster center it is closer;

[0129] (3) Statistically analyze the judgment results corresponding to each cluster center to obtain the probability value of the electromagnetic characteristic knowledge g of the scattering center of each target extremely narrow pulse echo data belonging to different cluster centers, and use this probability vector with a dimension of 1×k as the feature representation vector of the electromagnetic characteristic knowledge g of the target extremely narrow pulse echo data;

[0130] (4) Repeat (1) to (3) and concatenate the z types of feature representation vectors obtained to obtain a z × k target feature representation matrix for each target extremely narrow pulse echo data;

[0131] In the step 2, the method for preprocessing the target extremely narrow pulse echo data is:

[0132] In order to make full use of the amplitude and phase information of the target extremely narrow pulse echo data, the target extremely narrow pulse echo data X is divided into real and imaginary parts, and then the real and imaginary parts are spliced to form a data vector. Finally, the data vector is normalized to between 0 and 1 and represented by x:

[0133] x=norm([Re(X),Im(X)])

[0134] Wherein, Re(X) represents the operation of taking the real part of the target extremely narrow pulse echo data X, Im(X) represents the operation of taking the imaginary part of the target extremely narrow pulse echo data X, and norm represents the maximum value normalization operation;

[0135] The method for constructing a deep neural network is as follows: the deep neural network includes four network blocks, an average pooling layer, and a linear mapping layer. The first and third network blocks are the same and both contain two convolutional layers. The second and fourth network blocks have the same structure and both contain three convolutional layers. Each network block is constructed using a convolutional layer and a residual connection structure. The residual connection structure is proposed to solve the degradation problem of deep neural networks. It can make networks with more layers converge better and improve the learning ability and stability of the network. For the input x, the output O of the residual connection structure is expressed as:

[0136] O = relu(F(x)+x)

[0137] Among them, relu represents the "relu" nonlinear activation function, and F(x) represents the result obtained by the convolutional layer of the neural network;

[0138] Input the preprocessed data into the constructed deep neural network to obtain the target depth feature f d , f d The dimension is 1×d, where d is a set value and is a known value;

[0139] In step 3, the method of combining the target feature representation matrix with the target depth feature is:

[0140] 301. First, the target feature representation matrix is linearly mapped to z×d to obtain a representation closer to the target depth feature. Then, each row of the reduced target feature representation matrix is concatenated with the target depth feature, that is, each feature representation vector is combined with the target depth feature to obtain a feature combination matrix V with a dimension of z×2d.

[0141] 302. Input the feature combination matrix V obtained in step 301 into the attention mechanism, and redistribute weights to each feature combination vector in the feature combination matrix V. Feature combination vectors with large weights are considered to be able to better distinguish targets, so that the network pays more attention to more discriminative feature combination vectors.

[0142] The method for redistributing weights is as follows: the attention mechanism first uses an average pooling layer and two fully connected layers to learn a weight vector m corresponding to each feature combination vector, and then multiplies the weight vector m with the feature combination matrix V to obtain a weighted feature combination matrix V′;

[0143] The weight vector m is calculated as follows:

[0144] m=Pool(σ(W2δ(W2V)))

[0145] Among them, Pool() represents the average pooling operation, σ represents the "sigmoid" activation function, δ represents the "relu" activation function, and W2 and W1 are network parameters;

[0146] The weighted feature combination matrix V′ is:

[0147] V′=V·m;

[0148] In step 3, the method for predicting the target category based on the classification feature vector is:

[0149] Use two fully connected layers to predict the category of the target. Specifically, first flatten the weighted feature combination matrix V' into a one-dimensional vector, use the dimension P output by the first fully connected layer to reduce the dimension of the weighted feature combination matrix V', and use "relu" as the activation function to obtain the final classification feature vector f cls , the output dimension of the second fully connected layer is the number of target categories, and the activation function is "softmax". The category with the largest output probability is the predicted target category;

[0150] In step 4, the method for constructing the knowledge constraint loss function is:

[0151] 401. A contrast loss function is established based on the classification feature vector obtained in step 3 and the true category of the target to be detected. The contrast loss function is:

[0152]

[0153] Among them, Y represents the difference in the true categories of the two targets to be tested. Y = 1 means that the true categories of the two targets to be tested are the same, and Y = 0 means that the true categories of the two targets to be tested are different. cls1 and f cls2 Represents the classification feature vector of the two targets to be tested, N represents the number of targets to be tested, T is the set threshold, D represents the Euclidean distance between the classification feature vectors of the two targets to be tested, P represents the dimension of the classification feature vector, and the superscript p represents the pth feature in the classification feature vector;

[0154] When Y=1, the contrast loss function only has That is, when the true categories of the two targets to be tested are the same, the Euclidean distance of the classification feature vectors is large, which means that the deep neural network is not well trained;

[0155] When Y=0, the contrast loss function is That is, when the true categories of the two targets to be tested are different, the Euclidean distance of the classification feature vectors is large, which means that the deep neural network is well trained. When the Euclidean distance of the classification feature vectors exceeds T, the contrast loss function is set to 0;

[0156] 402. Using the prior knowledge of the true category hierarchy of the target to be measured, a logical relationship expression is established to improve the contrast loss function constructed in step 401 to obtain an improved contrast loss function;

[0157] The method of establishing a logical relationship expression based on the prior knowledge of the true category level of the target to be tested is:

[0158] Taking u1 and u2 as examples, we first calculate the true category hierarchy number C(u1) of target u1 and the true category hierarchy number C(u2) of target u2. Then we calculate the hierarchy number h of the two targets u1 and u2 whose true categories are the same. The logical relationship expression Y of the two targets is: c Expressed as:

[0159]

[0160] Assume that both targets u1 and u2 belong to the two categories of vehicles and cars, but the models are different. In this case, the number of true category levels is C(u1)=C(u2)=3, and the number of levels of the same true category of targets u1 and u2 is h=2, so we get Y c =2 / 3,

[0161] The improved contrast loss function is:

[0162]

[0163] 403. Combine the improved contrast loss function with the cross entropy loss function to construct the knowledge constraint loss function. The knowledge constraint loss function is:

[0164] L=γ1L C +γ2L CE

[0165] Among them, γ1 and γ2 are proportional coefficients, L C is the improved contrast loss function, L CE is the cross entropy loss function, y n is the true category of the nth target to be tested, is the predicted target category of the nth target to be tested.

[0166] In step 5, the method for training and updating the deep neural network parameters is:

[0167] During the training phase, the input of the deep learning network is the target containing the true category and the corresponding electromagnetic feature knowledge. For any target, the predicted category of the target is obtained through the network forward propagation, and the network loss is calculated using the knowledge constraint loss function L. Then, the network parameters are updated layer by layer through the Adam optimizer and the backpropagation algorithm to minimize the loss function and make the network converge, thus obtaining the trained and updated deep neural network.

[0168] In step 6, the method for performing target identification is:

[0169] The input of the deep learning network is the target of unknown true category and the corresponding electromagnetic feature knowledge. The target category is predicted through the trained and updated deep neural network. The category with the highest probability output by the last layer is the target category.

[0170] Example

[0171] This section illustrates the effectiveness of the invention by combining experiments with a publicly available simulation dataset. The dataset used is fully polarized target ultra-narrow pulse echo data obtained by the U.S. Air Force Laboratory using simulations of 10 civilian vehicles. These include the following: Jeep 93, Jeep 99, Mitsubishi, Toyota Avalon, Toyota Camry, Mazida MPV, Sentra, Toyota Tocoma, Honda Civic, and Maxima.

[0172] The center frequency of the dataset simulation settings is 9.6GHz, and the frequency modulation interval is 10.48MHz. The azimuth angle varies from 0 to 360°, with an angular interval of 0.0625°. Since the target has azimuth symmetry, this experiment selects data with an interval of 1° from 0 to 180° for the experiment, and uses a sub-dataset with a pitch angle of 30° that is closer to the actual application scenario. In addition, to be closer to the actual situation, 25dB Gaussian white noise is added to the target data, and 10 Monte Carlo simulations are performed, resulting in a total of 1,800 targets for each target. In the experiment, the training set and test set are divided into a ratio of 7:3, resulting in 12,600 training targets and 5,400 test targets.

[0173] In the experiment, the number of cluster centers selected by the bag-of-words model is set to k = 10, the depth feature dimension is d = 64, the feature dimension for classification is P = 32, and the threshold T in the contrast loss function is 5; the true category hierarchy prior knowledge of the 10 types of targets to be detected in the dataset is as follows Figure 1 shown.

[0174] The confusion matrix of the experimental results of the method proposed in this invention is as follows Figure 2As shown in the figure, it can be observed that the proposed method has good recognition effects on all 10 types of vehicles, among which the Jeep 93 has the highest recognition rate. The overall recognition accuracy of the proposed method is 93.24%, achieving good performance in the vehicle target recognition task.

[0175] In order to further evaluate the effectiveness of the method of the present invention, comparative experiments were conducted with three existing methods in the field of target extremely narrow pulse echo recognition. Among them, one method is a method that combines features with classifiers. The features utilize the feature representation matrix used for embedding the network in the present invention, and the classifier adopts the support vector machine (SVM) classifier. The other two methods are neural network methods, convolutional neural network (CNN) and long short-term memory network (LSTM), both of which have shown good results in the target extremely narrow pulse echo recognition task. The experimental results are shown in Table 1. It can be seen that the method of the present invention has the best recognition effect because it embeds target feature knowledge and establishes a knowledge constraint loss function. In addition, the effect of the neural network method is better than that of the method combining features with classifiers.

[0176] Table 1 Comparison of recognition results of different methods

[0177] method Accuracy Features + SVM classifier 76.19％ LSTM 88.72％ CNN 91.07％ Method of the present invention 93.24％

[0178] The present invention proposes a method for identifying extremely narrow pulse echoes of radar targets by combining prior knowledge with the assistance of deep neural networks. In order to improve the performance of deep neural networks, prior knowledge is introduced to assist in different stages of network training to achieve accurate classification of different targets. Specifically, the present invention effectively combines the knowledge of target electromagnetic scattering characteristics with the deep features extracted by the deep neural network to enhance the pertinence and discriminability of the deep neural network for feature representation learning. On the other hand, the present invention utilizes the prior knowledge of the real category hierarchical relationship of the target to be measured to construct a knowledge-constrained loss function to improve the recognition performance of the deep neural network. Experiments based on a fully polarized public simulation data set of 10 vehicle targets have verified the effectiveness of the method of the present invention.

[0179] In summary, the above are only preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for ultra-narrow pulse radar target recognition using a category and feature-assisted neural network, characterized in that The specific steps of this method include: Step 1: Extract the scattering center of the target extremely narrow pulse echo data and obtain the electromagnetic feature knowledge of the extracted scattering center. Then, use the extracted electromagnetic feature knowledge of the scattering center to generate a target feature representation matrix through the bag-of-words model; Step 2: Preprocess the target extremely narrow pulse echo data and use a deep neural network to extract the target depth features from the preprocessed data; Step 3: Combine the target feature representation matrix generated in step 1 with the target depth feature generated in step 2 to obtain a classification feature vector, and predict the target category based on the classification feature vector; Step 4: Construct a knowledge-constrained loss function based on the target category predicted in step 3 and the actual category level prior knowledge of the target to be detected; Step 5: Train and update the deep neural network parameters according to the knowledge constraint loss function constructed in step 4 to obtain the trained and updated deep neural network; Step 6: Perform target recognition based on the trained and updated deep neural network obtained in step 5 to obtain the target category, completing the ultra-narrow pulse radar target recognition using the category and feature-assisted neural network.

2. The method for ultra-narrow pulse radar target recognition using a category and feature-assisted neural network according to claim 1, characterized in that: In the step 1, the scattering centers of the target extremely narrow pulse echo data are extracted using an ordered statistical constant false alarm detection method, the number of extracted scattering centers is R, and any scattering center is represented by r, where r=1, 2, 3, ..., R; There are z kinds of electromagnetic characteristic knowledge of the scattering center r. When z=9, the electromagnetic characteristic knowledge are position proportion, intensity proportion, sparsity characteristic, scattering polarization entropy H, scattering polarization angle α, power matrix trace, polarization ratio, polarization correlation coefficient, and frequency domain polarization divergence.

3. The method for ultra-narrow pulse radar target recognition using a category and feature-assisted neural network according to claim 2, characterized in that: The position ratio indicates the position of the scattering center in the target area, and the length from the target starting point to the scattering center is the percentage LEN of the target area length. r Indicates as follows: Among them, len r Indicates the position of the scattering center r, len s Indicates the starting distance unit position of the target area, len e Indicates the end distance unit position of the target area; The intensity ratio indicates the ratio of the intensity of the scattering center to the total intensity of the target area, expressed as E r : Among them, e r represents the scattered power at the scattering center r; The sparsity feature represents the sparsity of the scattering center and the surrounding scattering centers. The variance of the distance between the rth scattering center and the left and right scattering centers is used as the sparsity feature Sp r ,as follows: Sp r =var(sp D1 ,sp D2 ) Among them, sp D1 is the distance between the scattering center r and the left scattering center, sp D2 is the distance between the scattering center r and the right scattering center, var(sp D1 ,sp D2 ) is sp D1 and sp D2 variance; The scattering polarization entropy H and the scattering polarization angle α are as follows: Where λ1 represents the maximum eigenvalue of the covariance matrix of the target extremely narrow pulse echo segment, λ2 represents the second largest eigenvalue of the covariance matrix of the target extremely narrow pulse echo segment, i = [1, 2], λ i represents the i-th eigenvalue of the covariance matrix of the target extremely narrow pulse echo segment; The power matrix trace is: Span r =|HH r | 2 +|VV r | 2 +|VH r | 2 +|HV r | 2 Among them, HH r is the amplitude of the scattering center r in the HH polarization mode; VV r is the amplitude of the scattering center r in HV polarization mode; VH r is the amplitude of the scattering center r in the VH polarization mode; HV r is the amplitude of the scattering center r in HV polarization mode; The polarization ratio and polarization correlation coefficient are: Polarization ratio Polarization correlation coefficient Where, <.> represents spatial averaging; superscript * indicates conjugation operation; j represents an imaginary number, represents the phase of the scattering center r in the HH polarization mode, represents the phase of the scattering center r in the VH polarization mode, To find the exponential function of the phase of the scattering center r in the HH polarization mode, To find the exponential function of the phase of the scattering center r in the VH polarization mode.

4. The method for ultra-narrow pulse radar target recognition using a category and feature-assisted neural network according to claim 3, characterized in that: In step 1, the method for generating a target feature representation matrix from the electromagnetic feature knowledge of the scattering center through the bag-of-words model is as follows: (1) Any electromagnetic feature knowledge is represented by g, where g = 1, 2, ..., z, and the electromagnetic feature knowledge g is divided into k different sets according to the differences between the values; (2) For each target extremely narrow pulse echo data, determine to which set the electromagnetic characteristic knowledge g of all scattering centers belongs in step (1); (3) Statistically analyzing the judgment results corresponding to each cluster center obtained in step (2) to obtain the probability value of the electromagnetic characteristic knowledge g of the scattering center of each target extremely narrow pulse echo data belonging to different cluster centers, and using this probability vector with a dimension of 1×k as the feature representation vector of the electromagnetic characteristic knowledge g of the target extremely narrow pulse echo data; (4) Repeat steps (1) to (3) to concatenate the z types of feature representation vectors obtained, and obtain a z×k target feature representation matrix for each target extremely narrow pulse echo data.

5. The method for ultra-narrow pulse radar target recognition using a category and feature-assisted neural network according to claim 4, characterized in that: In the step 2, the method for preprocessing the target extremely narrow pulse echo data is: The target extremely narrow pulse echo data X is divided into real and imaginary parts, and then the real and imaginary parts are spliced to form a data vector. Finally, the data vector is normalized to between 0 and 1 and represented by x: x=norm([Re(X),Im(X)]) Wherein, Re(X) represents the operation of taking the real part of the target extremely narrow pulse echo data X, Im(X) represents the operation of taking the imaginary part of the target extremely narrow pulse echo data X, and norm represents the maximum value normalization operation; In the step 2, the deep neural network includes 4 network blocks, also including an average pooling layer and a linear mapping layer. The first and third network blocks are the same and both contain 2 convolutional layers. The second and fourth network blocks have the same structure and both contain 3 convolutional layers. Each network block is constructed using a convolutional layer and a residual connection structure.

6. The method for ultra-narrow pulse radar target recognition using a category and feature-assisted neural network according to claim 5, characterized in that: In step 3, the method of combining the target feature representation matrix with the target depth feature is:

301. First, the target feature representation matrix is linearly mapped to z×d to obtain a representation closer to the target depth feature. Then, each row of the reduced target feature representation matrix is concatenated with the target depth feature to obtain a feature combination matrix V with a dimension of z×2d, where d is a set known value.

302. Input the feature combination matrix V obtained in step 301 into the attention mechanism, and reallocate the weights to each feature combination vector in the feature combination matrix V; The method for redistributing weights is as follows: the attention mechanism first uses an average pooling layer and two fully connected layers to learn a weight vector m corresponding to each feature combination vector, and then multiplies the weight vector m with the feature combination matrix V to obtain a weighted feature combination matrix V′; The weight vector m is calculated as follows: m=Pool(σ(W2δ(W1V))) Among them, Pool() represents the average pooling operation, σ represents the "sigmoid" activation function, δ represents the "relu" activation function, and W2 and W1 are network parameters; The weighted feature combination matrix V′ is: V′=V·m.

7. The method for ultra-narrow pulse radar target recognition using a category and feature-assisted neural network according to claim 6, characterized in that: In step 3, the method for predicting the target category based on the classification feature vector is: First, the weighted feature combination matrix V′ is flattened into a one-dimensional vector, and the dimension P of the first fully connected layer output is used to reduce the dimension of the weighted feature combination matrix V′. The activation function is "relu" to obtain the final classification feature vector f cls , the output dimension of the second fully connected layer is the number of target categories, and the activation function is "softmax". The category with the largest output probability is the predicted target category.

8. The method for ultra-narrow pulse radar target recognition using a category and feature-assisted neural network according to claim 7, characterized in that: In step 4, the method for constructing the knowledge constraint loss function is:

401. A contrast loss function is established based on the classification feature vector and the true category of the target to be detected. The contrast loss function is: Among them, Y represents the difference in the true categories of the two targets to be tested. Y = 1 means that the true categories of the two targets to be tested are the same, and Y = 0 means that the true categories of the two targets to be tested are different. cls1 and f cls2 Represents the classification feature vector of the two targets to be tested, N represents the number of targets to be tested, T is the set threshold, D represents the Euclidean distance between the classification feature vectors of the two targets to be tested, P represents the dimension of the classification feature vector, and the superscript p represents the pth feature in the classification feature vector; 402. Using the prior knowledge of the true category hierarchy of the target to be measured, a logical relationship expression is established to improve the contrast loss function constructed in step 401 to obtain an improved contrast loss function; The improved contrast loss function is:

403. Combine the improved contrast loss function with the cross entropy loss function to construct the knowledge constraint loss function. The knowledge constraint loss function is: L=γ1L C +γ2L CE Among them, γ1 and γ2 are proportional coefficients, L C is the improved contrast loss function, L CE is the cross entropy loss function, y n is the true category of the nth target to be tested, is the predicted target category of the nth target to be tested.

9. The method for ultra-narrow pulse radar target recognition using a category and feature-assisted neural network according to claim 8, characterized in that: In step 5, the method for training and updating the deep neural network parameters is: During the training phase, the input of the deep learning network is the target containing the real category and the corresponding electromagnetic feature knowledge. For any target, the predicted category of the target is obtained through the network forward propagation, and the knowledge constraint loss function L is used to calculate the network loss. Then, the network parameters are updated layer by layer through the Adam optimizer and backpropagation algorithm to minimize the loss function and make the network converge, thus obtaining the trained and updated deep neural network.

10. The method for ultra-narrow pulse radar target recognition using a category and feature-assisted neural network according to claim 9, characterized in that: In step 6, the method for performing target identification is: The input of the deep learning network is the target of unknown true category and the corresponding electromagnetic feature knowledge. The target category is predicted through the trained and updated deep neural network, and the category with the highest probability output by the last layer is the target category.

Images