A radar recognition method for unknown spatial cone targets

Abstract

The present invention discloses a radar recognition method for unknown spatial cone targets. The method comprises the following steps: obtaining a HRRP sequence to be tested; extracting weighted physical features and deep features of the HRRP sequence to be tested; fusing the weighted physical features and deep features to obtain fused features; inputting the fused features into a pre-trained prototype network to obtain multiple embedding vectors to be tested based on the prototype network; and performing radar recognition of unknown spatial cone targets based on the prototype vectors and the embedding vectors to be tested for each target type; the target types include spatial cone targets, light accompanying false targets, and heavy accompanying false targets. The radar recognition method for unknown spatial cone targets provided by the present invention solves the common problems of existing recognition methods, such as weak generalization ability of spatial cone targets and difficulty in zero-sample recognition.

Description

Technical Field

[0001] The invention belongs to the technical field of radar target recognition, and in particular relates to a radar recognition method for unknown spatial cone targets. Background Art

[0002] In modern military warfare, space cone targets have a long attack range and fast strike speed, and have strong strategic deterrence capabilities. However, during the flight of space cone targets, there are often a large number of accompanying false targets. Identifying space cone targets from the numerous accompanying false targets is of great significance to anti-missile defense and national security.

[0003] To effectively penetrate defenses, spatial cone targets and accompanying decoys typically maintain a small distance between them. Narrowband radars are limited by their bandwidth and have low range resolution. Multiple targets may reside within the same range cell of the low-resolution range profile echo, making it impossible to distinguish between different targets. Wideband radars, on the other hand, have a larger bandwidth and higher range resolution. Different targets reside in different range cells of the high-resolution range profile (HRRP) echo, facilitating target differentiation and further determining the specific categories of spatial cone targets and accompanying decoys. Furthermore, the wideband HRRP echo reflects the distribution of target scattering centers along the radar's line of sight, providing richer target details than narrowband echoes. Therefore, spatial cone target recognition based on wideband HRRP echoes is of great research significance and has garnered widespread attention.

[0004] Existing spatial cone target recognition methods based on radar HRRP sequences can be divided into two types: those based on manual feature extraction and those based on deep learning. The spatial cone target recognition method based on manual feature extraction first extracts statistical features such as the number of scattering points and energy entropy, or transform domain features such as the Doppler domain and time-frequency domain, from the radar target HRRP echo data. Based on this feature extraction, it uses shallow recognition models such as k-nearest neighbor and random forest to recognize spatial cone targets. The spatial cone target recognition method based on deep learning can approximate the complex function between the input radar target echo and the category label through multi-layer nonlinear transformations, completing end-to-end learning from input data to output category.

[0005] The patent with publication number CN115932778A, "A method for mid-course target recognition using HRRP sequence at low data rate", proposes a method for mid-course target recognition based on low data rate HRRP sequence. This method is based on the periodicity of the scattering center point in the HRRP sequence. Through time-frequency analysis, the energy distribution of the signal along the time dimension is obtained, and the target precession frequency is extracted. In addition, the number of target scattering points, target length, entropy and other features are extracted and spliced ​​with the precession frequency features. The fused features are classified using a random forest classifier. The document "HRRP image recognition of mid-course targets based on DCNN" proposes a spatial cone target HRRP image recognition method based on deep convolutional neural network (DCNN). The specific steps of this method are: first, the one-dimensional HRRP is converted into a 0-1 binary HRRP image using the HRRP visualization method; second, the 0-1 binary HRRP image features are extracted layer by layer using the DCNN combined with the L2 regularization constraint; third, the normalized exponential function is used based on the proposed features to realize the classification of spatial cone targets.

[0006] However, existing HRRP recognition methods for space cone targets are limited to classifying known space cone targets, light accompanying decoys, and heavy accompanying decoys that appeared during training. However, with the continuous optimization and updating of equipment, new, non-cooperative, unknown targets with parameters such as micro-motion and structure that differ from those of the known training targets often appear in actual combat. For unknown, variable-parameter targets not included in training, existing recognition methods suffer from weak generalization and difficulty in zero-shot recognition. Summary of the Invention

[0007] In order to solve the above problems existing in the prior art, the present invention provides a radar recognition method for unknown spatial cone targets.

[0008] The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0009] The present invention provides a radar recognition method for an unknown space cone target, the radar recognition method for an unknown space cone target comprising:

[0010] Obtain the HRRP sequence to be tested;

[0011] Extracting weighted physical features and deep features of the HRRP sequence to be measured;

[0012] fusing the weighted physical features and the deep features to obtain fused features;

[0013] Inputting the fused features into a pre-trained prototype network to obtain a plurality of embedding vectors to be tested based on the prototype network;

[0014] Unknown space cone target radar recognition is performed based on the prototype vector of each target type and the embedded vector to be tested; the target types include space cone targets, light accompanying false targets and heavy accompanying false targets.

[0015] Optionally, the method of extracting the weighted physical characteristics of the HRRP sequence to be measured includes:

[0016] Acquiring physical characteristics of the HRRP sequence to be measured;

[0017] The physical feature is subjected to nonlinear transformation and normalized exponential function to obtain a normalized physical feature weight vector;

[0018] The physical feature is weighted using the normalized physical feature weight vector to obtain a weighted physical feature.

[0019] Optionally, the physical characteristics include micro-motion parameters, structural parameters, maximum radial size characteristics and energy entropy characteristics.

[0020] Optionally, a method for obtaining the maximum radial dimension feature includes:

[0021] Acquiring energy distribution of the HRRP sequence to be measured;

[0022] determining a HRRP sequence signal search area based on the energy distribution;

[0023] determining a radial size of each HRRP echo in the HRRP sequence to be measured based on the HRRP sequence signal search area;

[0024] The largest radial dimension is selected as the maximum radial dimension feature.

[0025] Optionally, the radial size of each HRRP echo in the HRRP sequence to be measured is calculated as follows:

[0026]

[0027] Among them, L i represents the radial size of the i-th HRRP echo in the HRRP sequence to be measured; M represents the number of HRRP echoes in the HRRP sequence to be measured; ρ r Indicates the radar range resolution; Indicates the number of range units occupied by the radial size of the target in the i-th HRRP echo.

[0028] Optionally, the energy entropy feature is obtained by:

[0029] Acquiring energy distribution of the HRRP sequence to be measured;

[0030] Normalizing the energy distribution to obtain an energy probability distribution;

[0031] The information entropy of the energy probability distribution is calculated to obtain the energy entropy feature.

[0032] Optionally, the energy entropy feature is calculated as follows:

[0033]

[0034] Wherein, f7 represents the energy entropy characteristic; p r represents the rth element of the energy probability distribution, where r=1, ..., R; and R represents the number of range units of the HRRP echo in the HRRP sequence to be measured.

[0035] Optionally, the prototype vector is calculated as follows:

[0036] Obtain multiple sample HRRP sequences under each target type;

[0037] Extract sample weighted physical features and sample deep features of sample HRRP sequences under each target type;

[0038] For each target type of sample HRRP sequence, the sample weighted physical features and sample deep features of the sample HRRP sequence are fused to obtain the sample fusion features of the target type;

[0039] Inputting the sample fusion features under each target type into the prototype network to obtain multiple embedded sample vectors under each target type based on the prototype network;

[0040] For each target type, the average value of multiple embedded sample vectors under the target type is calculated to obtain the prototype vector under the target type.

[0041] The present invention provides a radar recognition method for unknown spatial cone targets, which characterizes the HRRP sequence to be tested by fusing weighted physical features with deep features, thereby solving the problems of incomplete feature representation and poor inter-class separability in the existing spatial cone target radar HRRP recognition methods that only use shallow features or pure data-driven deep features, and provides a good feature basis for subsequent recognition. In addition, the pre-trained prototype network used in the present invention can more accurately summarize class-general characteristics from known targets through pre-training, and has stronger generalization ability for similar heterogeneous targets, thereby solving the general problems of weak generalization ability and difficulty in zero-sample recognition for unknown new spatial cone targets that have not appeared in the training stage in the existing spatial cone target radar HRRP recognition methods, and improving the accuracy of unknown spatial cone target radar recognition.

[0042] The present invention will be further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 1 is a flow chart of a radar identification method for an unknown spatial cone target provided by an embodiment of the present invention;

[0044] Figure 2 Schematic diagram of energy distribution of the HRRP sequence to be tested provided by an embodiment of the present invention;

[0045] Figure 3 Schematic diagram of the number of range units occupied by a single HRRP echo target provided by an embodiment of the present invention;

[0046] Figure 4 Schematic diagram of the structure of the weight learning network provided by an embodiment of the present invention;

[0047] Figure 5 Schematic diagram of the structure of a deep feature extraction model based on a convolutional neural network provided by an embodiment of the present invention;

[0048] Figure 6 1 is a schematic diagram of the structure of an adaptive weighted network provided by an embodiment of the present invention;

[0049] Figure 7 This is a comparison diagram of the recognition accuracy confusion matrix of different recognition methods;

[0050] Figure 8 This is a t-SNE visualization diagram of the features proposed by different recognition methods and the learned classification interface;

[0051] Figure 9 It is a statistical diagram of the average value of the seven-dimensional physical feature weights of all test samples of various targets;

[0052] Figure 10This is a schematic diagram of corresponding samples correctly classified by a radar recognition method for unknown spatial cone targets provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0053] The present invention will be further described in detail below with reference to specific examples, but the embodiments of the present invention are not limited thereto.

[0054] In order to solve the technical problems of weak generalization ability and difficulty in zero-sample recognition in the existing space cone target radar HRRP recognition method, the embodiment of the present invention provides an unknown space cone target radar recognition method, see Figure 1 , Figure 1 FIG. 1 is a flow chart of a radar identification method for an unknown spatial cone target provided by an embodiment of the present invention, which specifically includes the following steps:

[0055] Step S101: Acquire the HRRP sequence to be tested.

[0056] In an embodiment of the present invention, a radar system can be used to detect a target. By transmitting a broadband signal and receiving echo signals reflected by the target, a series of signal processing steps are performed to extract a HRRP sequence to be detected that reflects the structural characteristics of the target. The HRRP sequence to be detected includes multiple HRRP echoes.

[0057] Step S102: extracting weighted physical features and deep features of the HRRP sequence to be measured.

[0058] In an embodiment of the present invention, the weighted physical characteristics of the HRRP sequence to be measured are obtained by weighting the physical characteristics of the HRRP sequence to be measured. First, the physical characteristics of the HRRP sequence to be measured are described. The physical characteristics of the HRRP sequence to be measured include micro-motion parameters, structural parameters, maximum radial dimension characteristics, and energy entropy characteristics. The micro-motion parameters include precession frequency f1 and precession angle f2. The structural parameters include cone height f3, cone base radius f4, and cone generatrix length f5.

[0059] The micro-motion parameters and structural parameters can be obtained by extracting the scattering point macro-distance curve from the HRRP sequence diagram and then solving the extracted scattering point macro-distance curve and the scattering point theory macro-distance equation. For details, please refer to the existing technology and will not be repeated here.

[0060] In one implementation, the maximum radial dimension feature is obtained by:

[0061] Obtaining the energy distribution of the HRRP sequence to be measured;

[0062] Determine the HRRP sequence signal search area based on energy distribution;

[0063] determining the radial size of each HRRP echo in the HRRP sequence to be measured based on the HRRP sequence signal search area;

[0064] The largest radial dimension is selected as the maximum radial dimension feature.

[0065] In the embodiment of the present invention, firstly, all HRRP echoes in the HRRP sequence to be measured are accumulated to obtain the energy distribution of the HRRP sequence to be measured. Figure 2 , Figure 2 This is a schematic diagram of the energy distribution of the HRRP sequence to be tested provided by an embodiment of the present invention. Then, the mean value of the energy distribution of the HRRP sequence to be tested is used as the threshold, and the HRRP sequence area to be tested within the distance units on the left and right sides that are greater than the threshold is used as the HRRP sequence signal search area. For each HRRP echo in the HRRP sequence signal search area, all peak points in the HRRP echo whose amplitude is greater than the mean value of the echo amplitude and greater than the echo amplitude of two adjacent distance units are searched, and the distance unit indexes of all peak points are determined. The maximum index value minus the minimum index value is the number of distance units occupied by the target radial size in the HRRP echo. See also Figure 3 , Figure 3 This is a schematic diagram of the number of distance units occupied by a single HRRP echo target provided by an embodiment of the present invention. Taking the 32nd HRRP echo as an example, peak point 1 corresponds to the minimum index k1, peak point 2 corresponds to the maximum index k2, and the number of distance units occupied by the radial size of the target in the 32nd HRRP echo is w. opt =k2-k1.

[0066] The radial size of the target in the HRRP echo is further calculated:

[0067]

[0068] Among them, L i represents the radial size of the i-th HRRP echo in the HRRP sequence to be measured, M represents the number of HRRP echoes in the HRRP sequence to be measured, and ρ r Indicates the radar range resolution.

[0069] Finally, the maximum radial dimension f6 of the HRRP sequence to be measured is the maximum radial length of all HRRP echoes in the sequence.

[0070] In one implementation, the energy entropy feature is obtained by:

[0071] Obtaining the energy distribution of the HRRP sequence to be measured;

[0072] Normalize the energy distribution to obtain the energy probability distribution;

[0073] The information entropy of the energy probability distribution is calculated to obtain the energy entropy characteristics.

[0074] In this embodiment of the present invention, the energy distribution of the HRRP sequence to be measured is normalized to obtain the energy probability distribution p, and the information entropy of the energy probability distribution p is calculated to obtain the HRRP sequence energy entropy feature f7. The calculation method of f7 is:

[0075]

[0076] Among them, p r represents the rth element of the energy probability distribution p, where r=1,…,R; R represents the number of range units of the HRRP echo in the HRRP sequence to be measured.

[0077] Finally, the seven-dimensional physical features mentioned above are concatenated to obtain the physical feature f1 = [f1, f2, …, f7] of the HRRP sequence to be tested.

[0078] In one implementation, extracting the weighted physical characteristics of the HRRP sequence to be measured includes:

[0079] Obtaining physical characteristics of the HRRP sequence to be tested;

[0080] The physical features are subjected to nonlinear transformation and normalized exponential function to obtain the normalized physical feature weight vector;

[0081] The physical features are weighted using the normalized physical feature weight vector to obtain weighted physical features.

[0082] In the embodiment of the present invention, the physical feature f1 = [f1, f2, ..., f7] is first input into the weight learning network g composed of a multi-layer fully connected network, see Figure 4 , Figure 4 It is a structural diagram of the weight learning network provided by an embodiment of the present invention. After two layers of nonlinear transformation and normalized exponential function, the normalized physical feature weight vector ω = [ω1, ω2, ..., ω7] is obtained, and then the physical features are weighted to obtain the weighted physical features f1′ = [ω1f1, ω2f2, ..., ω7f7].

[0083] In an embodiment of the present invention, the nonlinear transformation is expressed as:

[0084] f l+1 =R(a l+1 f l +b l+1 );

[0085] Where R(·) represents the linear rectification function; f l represents the input of the lth layer nonlinear transformation; f l+1Represents the output of the lth layer nonlinear transformation, l = 1, 2, and f 1 =f1;a l+1 represents the weight of the nonlinear transformation of the lth layer; b l+1 Represents the bias of the nonlinear transformation of the lth layer.

[0086] In an embodiment of the present invention, the normalized exponential function is expressed as:

[0087]

[0088] Among them, x i represents the i-th element of the vector obtained by two layers of nonlinear transformation of f1; x k represents the kth element of the vector obtained by two layers of nonlinear transformation of f1; ω i Represents the i-th element of the normalized physical feature weight vector.

[0089] In an embodiment of the present invention, the extraction of deep features of the HRRP sequence to be measured can be achieved by using a deep feature extraction model based on a convolutional neural network (CNN), as follows:

[0090] The HRRP sequence to be tested is normalized, and then the normalized HRRP sequence to be tested is convolved using multiple convolution kernels to obtain a multi-channel feature map; then the multi-channel feature map is subjected to a channel attention operation to obtain a weighted feature map, see Figure 5 , Figure 5 This is a structural diagram of a deep feature extraction model based on a convolutional neural network provided by an embodiment of the present invention. The specific process is: a multi-channel feature map is subjected to global average pooling to obtain a feature vector, and then the feature vector is subjected to two layers of nonlinear transformation to obtain a feature map weight vector, and then the multi-channel feature map is weighted according to the feature map weight vector to obtain a weighted feature map; finally, the weighted feature map is subjected to activation and pooling operations to obtain the first layer of features; the first layer of features is subjected to a set of "convolution-channel attention-activation-pooling" operations in sequence to obtain the second layer of features; the second layer of features is subjected to a set of "convolution-channel attention-activation-pooling" operations to obtain the third layer of features, and the third layer of features is input into a fully connected network with a three-layer nonlinear mapping structure for feature dimensionality reduction to obtain the deep features f2 of the HRRP sequence to be measured. The convolutional neural network in the embodiment of the present invention uses a convolution kernel size of 3×3, a stride of 1×1 for the convolution operation, and a pooling window size of 2×2.

[0091] Step S103: fusing weighted physical features and deep features to obtain fused features.

[0092] In the embodiment of the present invention, the weighted physical feature f1′ and the deep feature f2 are first concatenated and input into the adaptive weighted network g′ composed of a multi-layer fully connected network. Figure 6 , Figure 6 This is a structural diagram of the adaptive weighted network provided by an embodiment of the present invention. After two layers of nonlinear transformation and normalized exponential function, a normalized two-dimensional weight vector [λ1,λ2] is obtained. Then, the weighted physical features and deep features of the HRRP sequence to be tested are weightedly spliced ​​to obtain the fusion feature d = [λ1f1′,λ2f2].

[0093] In step S104, the fused features are input into the pre-trained prototype network to obtain a plurality of embedding vectors to be tested based on the prototype network.

[0094] In the embodiment of the present invention, the prototype network has a two-layer nonlinear mapping structure. The prototype network is a model for few-shot learning that performs classification by learning the "prototype" of each category, that is, the average representation of samples of that category in the feature space. Specifically, the prototype network is trained as follows:

[0095] Obtain the fusion features of the known space cone target, the light accompanying decoy target, and the heavy accompanying decoy target. The method of obtaining the fusion features of these three types of targets can be found in the above steps and will not be repeated here.

[0096] Then, the fusion features of the known space cone target, light accompanying false target, and heavy accompanying false target are input into the prototype network to obtain multiple embedded sample vectors for each target type. Then, for each target type, the average value of the multiple embedded sample vectors under this target type is calculated to obtain the prototype vector under this target type, which is recorded as the prototype vector c under the known space cone target. 1 , prototype vector c under light accompanying false target 2 , prototype vector c under heavy accompanying decoy target 3 .

[0097] Assume that after feature fusion, the embedded sample vector corresponding to the HRRP sequence of the i-th sample of the spatial cone target is The embedded sample vector corresponding to the j-th sample HRRP sequence of the light accompanying false target is The embedded sample vector corresponding to the kth sample HRRP sequence of the heavy accompanying false target is Then the loss function L of the prototype network is proto It can be expressed as:

[0098]

[0099] Where N1 represents the total number of sample HRRP sequences of spatial cone targets; N2 represents the total number of sample HRRP sequences of light accompanying false targets; N3 represents the total number of sample HRRP sequences of heavy accompanying false targets; dis(x,y) represents the cosine similarity between vectors x and y, which is calculated as follows:

[0100]

[0101] Among them, xy T represents the inner product of two row vectors, the superscript T represents the transpose operation of the vector; |x||y| represents the module multiplication of the vectors; vectors x and y represent any two vectors of the same latitude; during the training phase, the convolutional neural network, physical feature weight learning network, adaptive weighted network, and prototype network are optimized simultaneously, and the optimal parameters of the network are obtained by minimizing the loss function of the prototype network.

[0102] At this point, a trained prototype network is obtained, and then the fusion features of the HRRP sequence to be tested are input into the pre-trained prototype network, and multiple embedding vectors to be tested can be obtained based on the prototype network.

[0103] Step S105 , performing radar recognition of unknown space cone targets based on the prototype vectors and the embedded vectors to be tested of each target type; target types include space cone targets, light accompanying false targets, and heavy accompanying false targets.

[0104] In this embodiment of the present invention, the aforementioned prototype vector for each target type refers to a representative vector in the feature space for each target type, representing the central feature of that target type. The test embedding vector refers to the embedding vector generated by mapping the test HRRP sequence to the feature space through feature extraction, reflecting the feature representation of the test HRRP sequence.

[0105] Target identification is achieved by calculating the distance between the embedding vector to be tested and the prototype vectors of each known category, finding the category with the closest distance. Specifically, by measuring the similarity between each embedding vector to be tested and each prototype, the HRRP sequence to be tested is identified as the target type with the greatest similarity. The number of instances of the HRRP sequence to be tested assigned to each known category is counted, and the target of the HRRP sequence to be tested is identified as the category corresponding to the known category with the greatest number of assigned instances. The similarity calculation method is the same as the cosine similarity calculation formula described above and will not be repeated here.

[0106] In an embodiment of the present invention, the HRRP sequence to be tested is characterized by fusing weighted physical features with deep features, thereby solving the problems of incomplete feature representation and poor inter-class separability in the existing spatial cone target radar HRRP recognition method that only uses shallow features or purely data-driven deep features, and providing a good feature basis for subsequent recognition; in addition, the pre-trained prototype network used in the embodiment of the present invention can more accurately summarize class general characteristics from known targets through pre-training, and has stronger generalization ability for similar heterogeneous targets, thereby solving the general problems of weak generalization ability and difficulty in zero-sample recognition for unknown new spatial cone targets that have not appeared in the training stage in the existing spatial cone target radar HRRP recognition method, and improving the accuracy of unknown spatial cone target radar recognition.

[0107] A simulation experiment using a radar recognition method for unknown spatial conical targets provided by an embodiment of the present invention is as follows: The radar echoes for known training targets and unknown test targets used in this simulation experiment are HRRP sequence samples of three categories: "spatial conical targets," "light accompanying false targets," and "heavy accompanying false targets," obtained using electromagnetic simulation. The radar parameters are set as follows: operating frequency 10 GHz, operating bandwidth 1.3 GHz, signal frequency range 9.35 GHz to 10.65 GHz, frequency interval 0.01 GHz, radar line of sight range 52° to 56°, repetition rate 100 Hz, sampling time 2 s, Gaussian white noise added to the echo signal, and signal-to-noise ratio 20 dB. The micro-motion and structural parameter settings for various known training targets are shown in Table 1. In the simulation experiment, the parameter settings for targets of the same type are not fixed; multiple values ​​are set for each micro-motion and structural parameter. Considering that the height of a conical target is generally greater than its diameter, different parameter settings are combined to form multiple targets of the same type. In the experiment, each of the three known classes contains 64 targets, each target contains 25 HRRP sequence samples, and the three known classes have a total of 4800 samples.

[0108] Table 1

[0109]

[0110] The micro-motion and structural parameter settings of the unknown target are different from those of the known target. The micro-motion and structural parameter settings of various unknown target tests are shown in Table 2. In the experiment, the three unknown classes each contain 30 targets, each target contains 20 HRRP sequence samples, and the three known classes have a total of 1800 samples.

[0111] Table 2

[0112]

[0113] After the training process obtains the optimal parameters of each network by minimizing the prototype contrast loss, the average of all embedding vectors of each type output by the prototype network is calculated to obtain various prototype vectors. In the testing phase, the three types of HRRP sequence samples of unknown targets to be tested are generated based on the above simulation conditions, and are input into the trained recognition network. After multi-dimensional physical feature extraction and deep feature extraction, physical feature selection, feature fusion and prototype network, the embedding vectors of the HRRP samples to be tested are obtained. The target type is determined by measuring the similarity between the embedding vector to be tested and various prototypes. The embodiment of the present invention defines the calculation formula for the average recognition accuracy of the three types of targets as follows:

[0114]

[0115] Among them, n i Indicates the number of correctly identified samples of the i-th category, N i Represents the total number of samples of the i-th category target.

[0116] According to the above recognition accuracy calculation formula, the average recognition rate of the three types of targets of the method proposed in the embodiment of the present invention is calculated, and compared with the recognition rates of four comparison methods: a method using only physical features and adopting a prototype network as a classifier (method 1), a method using only weighted physical features and adopting a prototype network as a classifier (method 2), a method using only deep features and adopting a prototype network as a classifier (method 3), and a method using fusion features based on an adaptive weighted network and a traditional normalized index classifier (method 4). Among them, the normalized index classifier in method 4 refers to a classification score prediction after performing a layer of nonlinear transformation on the fusion features and then passing a normalized exponential function. The different feature comparison methods in the simulation experiment are shown in Table 3:

[0117] Table 3

[0118]

[0119]

[0120] The above methods were compared with the test set samples of three types of targets: unknown space cone target (UA), unknown light accompanying false target (UB), and unknown heavy accompanying false target (UC). The average recognition accuracy of the three types of targets of each method is shown in Table 3. In order to quantitatively analyze the separability between different targets, the confusion matrix of different methods is given below. Figure 7 As shown, Figure 7 This is a comparison diagram of the confusion matrix of recognition accuracy of different recognition methods. Figure 7 (a) is a schematic diagram of the recognition accuracy confusion matrix of method 1. Figure 7 (b) is a schematic diagram of the recognition accuracy confusion matrix of method 2. Figure 7(c) is a schematic diagram of the recognition accuracy confusion matrix of method 3. Figure 7 (d) is a schematic diagram of the recognition accuracy confusion matrix of method 4. Figure 7 (e) is a schematic diagram of the recognition accuracy confusion matrix of the unknown spatial cone target radar recognition method provided by an embodiment of the present invention. Figure 7 The vertical axis of the confusion matrix is ​​the true label of the three types of test samples, "UA", "UB", and "UC", respectively representing "space cone target", "light accompanying false target", and "heavy accompanying false target", and the horizontal axis is the category predicted by the recognition model.

[0121] From Table 3 and Figure 7 It can be seen from the results that: Method 1 and Method 2 are easy to distinguish heavy accompanying false targets, and their misclassified samples are mainly concentrated between spatial cone targets and light accompanying false targets. The recognition rate of Method 2 is improved by 1.83% compared with Method 1, which verifies the effectiveness of physical feature selection; Method 3 is more separable from spatial cone targets and light accompanying false targets than Method 2, and its misclassified samples are concentrated between light accompanying false targets and heavy accompanying false targets; the method proposed in the embodiment of the present invention combines the advantages of physical features and deep features, and can significantly improve the overall recognition performance of spatial cone targets, light accompanying false targets and heavy accompanying false targets. The recognition accuracy is improved by 6.89% and 11.17% compared with the comparison methods 2 and 3, respectively, which verifies the effectiveness of the feature fusion method based on adaptive weighted network proposed in the embodiment of the present invention; the recognition rate is improved by 3.01% compared with Method 4. The method proposed in the present invention adopts the prototype network as the classifier, which has stronger generalization for unknown targets.

[0122] See also Figure 8 , Figure 8 This is a t-SNE visualization diagram of the features proposed by different recognition methods and the classification interface learned. Figure 8 The t-SNE visualization of the features proposed by different methods and the learned classification interface in Table 3 is given to further qualitatively illustrate the recognition performance of different methods, where t-SNE is a nonlinear dimensionality reduction technique. Figure 8 (a) is the t-SNE visualization diagram of method 1. Figure 8 (b) is the t-SNE visualization diagram of method 2. Figure 8 (c) in the figure is the t-SNE visualization diagram of method 3. Figure 8 (d) is the t-SNE visualization diagram of method 4. Figure 8 (e) is a t-SNE visualization diagram of the unknown spatial cone target radar recognition method provided by an embodiment of the present invention. Figure 8In the figure, the red, green and black dots represent the features of unknown space cone targets, light accompanying false targets and heavy accompanying false targets respectively. The blue, pink and yellow five-pointed stars represent the space cone target prototypes, light accompanying false targets prototypes and heavy accompanying false targets prototypes learned in the training phase respectively.

[0123] from Figure 8 It can be seen that the unknown heavy accompanying false target features proposed by Method 1 and Method 2 are significantly different from the other two types of unknown target features compared with Method 3, which further illustrates the advantages of physical features in identifying heavy accompanying false targets; the aliasing degree of the unknown spatial cone target and the unknown light accompanying false target features learned by Method 3 is weaker than that of Method 1 and Method 2, which further illustrates the advantages of deep features in identifying spatial cone targets and light accompanying false targets. The feature separability of the three types of targets proposed by the method of the present invention is significantly enhanced, which once again verifies the effectiveness of the feature fusion method based on adaptive weighted network proposed by the present invention in extracting separable features. In addition, by comparison Figure 8 (d) and Figure 8 As can be seen from (e) in the figure, for the same fusion features, the classification interface learned by the classifier used in method 4 is significantly less accurate than the classification interface learned by the classifier used in the present invention, which further illustrates the superiority of the prototype network-based radar target recognition framework proposed in the present invention in recognizing unknown targets that have not appeared in the training phase.

[0124] To verify the rationality of the physical feature weight learning network, see Figure 9 , Figure 9 It is a statistical diagram of the average value of the seven-dimensional physical feature weights of all test samples of various targets. Figure 9 (a) is a statistical diagram of the average value of the seven-dimensional physical feature weights of all test samples of the spatial cone target. Figure 9 (b) is a statistical diagram of the average value of the seven-dimensional physical feature weights of all test samples of light accompanying decoy targets. Figure 9 (c) is a statistical diagram of the average value of the seven-dimensional physical feature weights of all test samples of heavy accompanying decoy targets. Figure 9 It can be seen that the weights of the size and structure-related features of heavy accompanying decoy targets are larger, while the weights of the micro-motion features of space cone targets and light accompanying decoy targets are larger. This shows that the physical features mainly rely on size and structure features to distinguish heavy accompanying decoy targets, and rely on micro-motion features to distinguish space cone targets and light accompanying decoy targets. This is consistent with the setting in Table 2 that the unknown heavy accompanying decoy targets and space cone targets / light accompanying decoy targets have large differences in structural parameters, and the space cone targets and light accompanying decoy targets have small differences in structural parameters but relatively large differences in micro-motion parameters, which verifies the rationality of the physical feature weight learning network.

[0125] To further verify the rationality of deep features in alleviating the misclassification of space cone targets and light accompanying false targets, see Table 4 and Figure 10 ,Table 4 is an example table of misclassified sample parameters for method 2, Figure 10 This is a schematic diagram of samples corresponding to the correct classification of an unknown spatial cone target radar recognition method provided by an embodiment of the present invention. Figure 10 (a) is a schematic diagram of the sample corresponding to the correct classification of the spatial cone target in Example 1. Figure 10 (b) is a schematic diagram of the correct classification of the light accompanying false target in Example 1. Figure 10 (c) in the figure is a schematic diagram of the corresponding sample of the correct classification of the spatial cone target in Example 2. Figure 10 (d) in Example 2 is a schematic diagram of the correct classification of the light accompanying false target, Table 4 and Figure 10 The corresponding samples correctly classified by the proposed method are given respectively. The parameter setting examples of the space cone target and the light accompanying false target that can be correctly distinguished by the proposed method in the embodiment of the present invention but misclassified by method 2 are given respectively, as well as the corresponding HRRP sequence examples. As can be seen from Table 4, since method 2 mainly relies on the micro-motion parameter features to distinguish the space cone target and the light accompanying false target, when the micro-motion parameter difference between the two is small, method 2 is easy to make an identification error. However, small micro-motion parameter differences may produce obvious changes in the HRRP sequence data, such as Figure 10 The number of sinusoidal cycles shown is significantly different. Therefore, it is meaningful to effectively distinguish between spatial cone targets and light accompanying false targets by fusing the deep features of HRRP echo data.

[0126] Table 4

[0127]

[0128] It should be noted that the terms "first," "second," and the like are used to distinguish similar objects and are not necessarily used to describe a particular order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate, so that the embodiments of the present invention described herein can be implemented in sequences other than those illustrated or described herein. The embodiments described in the following exemplary embodiments are not intended to represent all embodiments consistent with the present invention.

[0129] In the description of this specification, the reference terms "one embodiment," "some embodiments," "example," "specific example," or "some examples" mean that the specific features or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described can be combined in any suitable manner in any one or more embodiments or examples. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification.

[0130] Although the present invention is described herein in conjunction with various embodiments, in the process of implementing the claimed invention, those skilled in the art can understand and implement other variations of the disclosed embodiments by viewing the drawings and the disclosed content. In the description of the present invention, the word "comprising" does not exclude other components or steps, "one" or "a" does not exclude multiple situations, and "multiple" means two or more, unless otherwise clearly and specifically defined. In addition, certain measures are recorded in different embodiments, but this does not mean that these measures cannot be combined to produce good results.

[0131] The above is a further detailed description of the present invention in conjunction with specific preferred embodiments, and the specific implementation of the present invention should not be considered to be limited to these descriptions. For those skilled in the art of the present invention, without departing from the concept of the present invention, several simple deductions or substitutions can be made, which should be considered to fall within the scope of protection of the present invention.

Claims

1. A radar recognition method for unknown spatial cone targets, characterized in that: The unknown space cone target radar recognition method includes: Obtain the HRRP sequence to be tested; Extracting weighted physical features and deep features of the HRRP sequence to be measured; fusing the weighted physical features and the deep features to obtain fused features; Inputting the fused features into a pre-trained prototype network to obtain a plurality of embedding vectors to be tested based on the prototype network; Performing radar recognition of unknown space cone targets based on prototype vectors of each target type and the embedded vector to be tested; the target types include space cone targets, light accompanying false targets, and heavy accompanying false targets; The method of extracting the weighted physical characteristics of the HRRP sequence to be measured includes: Acquiring physical characteristics of the HRRP sequence to be measured; The physical feature is subjected to nonlinear transformation and normalized exponential function to obtain a normalized physical feature weight vector; weighting the physical feature using the normalized physical feature weight vector to obtain a weighted physical feature; The physical characteristics include micro-motion parameters, structural parameters, maximum radial size characteristics and energy entropy characteristics; The method for obtaining the maximum radial dimension feature includes: Acquiring energy distribution of the HRRP sequence to be measured; determining a HRRP sequence signal search area based on the energy distribution; determining a radial size of each HRRP echo in the HRRP sequence to be measured based on the HRRP sequence signal search area; Selecting the largest radial dimension as the maximum radial dimension feature; The radial size of each HRRP echo in the HRRP sequence to be measured is calculated as follows: ； in, Indicates the first The radial size of each HRRP echo; represents the number of HRRP echoes in the HRRP sequence to be measured; Indicates the radar range resolution; Indicates the The number of range units occupied by the radial dimension of the target in each HRRP echo; The energy entropy characteristics are obtained by: Acquiring energy distribution of the HRRP sequence to be measured; Normalizing the energy distribution to obtain an energy probability distribution; Calculating the information entropy of the energy probability distribution to obtain the energy entropy feature; The energy entropy characteristic is calculated as follows: ； in, represents the energy entropy characteristic; The energy probability distribution is represented by elements, ; Indicates the number of range units of the HRRP echo in the HRRP sequence to be measured.

2. The unknown space cone target radar recognition method according to claim 1, characterized in that: The calculation method of the prototype vector includes: Obtain multiple sample HRRP sequences under each target type; Extract sample weighted physical features and sample deep features of sample HRRP sequences under each target type; For each target type of sample HRRP sequence, the sample weighted physical features and sample deep features of the sample HRRP sequence are fused to obtain the sample fusion features of the target type; Inputting the sample fusion features under each target type into the prototype network to obtain multiple embedded sample vectors under each target type based on the prototype network; For each target type, the average value of multiple embedded sample vectors under the target type is calculated to obtain the prototype vector under the target type.

Images