Mask-assisted network robust radar target recognition method under time resource constraints

Abstract

The present invention discloses a mask-assisted network robust radar target recognition method under time resource constraints, which relates to the field of radar signal processing technology. The method comprises the following steps: obtaining a test echo signal; wherein the test echo signal includes mask information; inputting the test echo signal and the mask information of the test echo signal into a trained mask information-assisted network to obtain a recognition result; wherein the mask information-assisted network includes an improved Transformer encoder, a recognition module, and a mask matrix generation module, wherein the mask matrix generation module is used to generate a mask matrix; the improved Transformer encoder extracts features from the target side echo signal in combination with the mask matrix, and outputs the feature-extracted echo signal; and the recognition module recognizes the feature-extracted echo signal and outputs the recognition result. The present invention has good recognition robustness performance.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar signal processing, and in particular relates to a mask-assisted network robust radar target recognition method under time resource constraints. Background Art

[0002] Radar target recognition involves using a target's radar echo signal to determine its type. Ideally, a radar observes a single target, with a long, continuous dwell time. However, in practice, due to the need to detect and track multiple targets across the entire airspace, time resources are limited. After a brief observation of a single target, the beam will be directed toward other targets in the airspace and then redirected back toward the target. This results in discontinuities in the echo signal for each target, with gaps in some areas of the echo. These gaps do not contain structural information about the target, so directly feeding the echo into the recognition model will affect the model's recognition performance. Consequently, using traditional recognition methods in these situations will result in reduced target recognition accuracy.

[0003] In the prior art, the radar multiple short-dwell observation fusion classification method mainly involves voting and fusing the classification results of the radar single short-dwell observation samples. The voting method is a simple fusion method, and the basic principle is "the minority obeys the majority". However, since the echo of a single short-dwell observation contains a small amount of target information, the recognition result after the fusion of multiple short-dwell observations is not robust. When the number of votes is the same, the voting result cannot be determined. Therefore, the result is not robust when the recognition results of a single short-dwell radar observation sample are fused by voting. Li Haoming et al. proposed a radar target recognition method based on multi-frequency domain feature fusion under short-dwell conditions in the patent "Radar target recognition method based on multi-frequency domain feature fusion under short-dwell conditions". This method combines the echoes of multiple short-dwell observations of the radar on the target, extracts the corresponding features and sends them to the SVM classifier for recognition. However, this method requires that the short-dwell time must be the same each time, which is difficult to meet in actual application scenarios. On the one hand, the signal-to-noise ratio of targets at different distances is different. Targets with high signal-to-noise ratios are typically allocated less time, resulting in shorter single dwell times; targets with low signal-to-noise ratios are typically allocated more time, resulting in longer single dwell times. Furthermore, as the number of airspace targets increases, the dwell time allocated to each aircraft decreases accordingly; as the number of airspace targets decreases, the dwell time allocated to each aircraft increases accordingly. This shift in dwell time between training and testing can lead to different sample dimensions, causing the model's classification and recognition performance to degrade or even fail.

[0004] Therefore, there is an urgent need to improve the defects in the prior art. Summary of the Invention

[0005] In order to solve the above problems existing in the prior art, the present invention provides a mask-assisted network robust radar target recognition method under time resource constraints. The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0006] In a first aspect, the present invention provides a mask-assisted network robust radar target recognition method under time resource constraints, comprising:

[0007] Acquire an echo signal to be measured; wherein the echo signal to be measured includes mask information;

[0008] The echo signal to be measured and the mask information of the echo signal to be measured are input into the trained mask information auxiliary network to obtain the recognition result; wherein, the mask information auxiliary network includes an improved Transformer encoder, a recognition module and a mask matrix generation module, the mask matrix generation module generates a mask matrix according to the mask information of the echo signal to be measured, the improved Transformer encoder combines the mask matrix to extract features of the echo signal on the side, and outputs the echo signal after feature extraction; the recognition module recognizes the echo signal after feature extraction and outputs the recognition result.

[0009] Beneficial effects of the present invention:

[0010] The present invention provides a mask-assisted network robust radar target recognition method under time resource constraints, which inputs the echo signal to be measured and the mask information of the echo signal to be measured into a trained mask information auxiliary network to obtain a recognition result. The mask information auxiliary network is obtained by training based on a network training data set. The mask information auxiliary network includes a Transformer encoder, a recognition module and a mask matrix generation module; wherein, the echo signal is reconstructed according to the input echo signal to be measured and the mask information contained in the echo signal to be measured, so that the parts of the new echo signal containing structural information are spliced ​​together; the mask matrix generation module generates a corresponding mask matrix according to the mask information of the input echo signal to be measured. array; the improved Transformer encoder is combined with the mask matrix to extract features of the input echo signal to be tested, and the Transformer encoder output corresponding to the echo containing the target structure information is used as the input of the recognition module to finally obtain the target recognition result; in this way, the mask information of the missing part of the radar signal echo is used to assist the network to focus on the part of the echo information containing the target structure, and ignore the part of the echo information not containing the target structure. According to the mask information, the missing echo value is prevented from interfering with the target recognition process, thereby improving the recognition performance of the model for missing signals. When the radar echo signal is missing, it still has a good recognition effect and has good recognition robustness.

[0011] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 This is a flow chart of a mask-assisted network robust radar target recognition method under time resource constraints provided by an embodiment of the present invention;

[0013] Figure 2 1 is a schematic diagram of the structure of a mask information auxiliary network provided by an embodiment of the present invention;

[0014] Figure 3 1 is a diagram of simulation experiment results provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0015] 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.

[0016] See Figure 1 and Figure 2 As shown, Figure 1 This is a flow chart of a mask-assisted network robust radar target recognition method under time resource constraints provided by an embodiment of the present invention. Figure 2 This is a structural diagram of a mask information assisted network provided by an embodiment of the present invention. The present invention provides a mask assisted network robust radar target recognition method under time resource constraints, including:

[0017] S101, obtaining an echo signal to be measured; wherein the echo signal to be measured includes mask information;

[0018] S102. Input the echo signal to be measured and the mask information of the echo signal to be measured into the trained mask information auxiliary network to obtain a recognition result; wherein, the mask information auxiliary network includes an improved Transformer encoder, a recognition module and a mask matrix generation module, the mask matrix generation module generates a mask matrix according to the mask information of the echo signal to be measured, the improved Transformer encoder combines the mask matrix to extract features of the echo signal to the side, and outputs the echo signal after feature extraction; the recognition module recognizes the echo signal after feature extraction and outputs the recognition result.

[0019] Specifically, the mask-assisted network robust radar target recognition method under time resource constraints provided in this embodiment inputs the echo signal to be measured and the mask information of the echo signal to be measured into the trained mask information auxiliary network to obtain a recognition result. The mask information auxiliary network is trained based on a network training data set. The mask information auxiliary network includes a Transformer encoder, a recognition module and a mask matrix generation module; wherein, the echo signal is reconstructed according to the input echo signal to be measured and the mask information contained in the echo signal to be measured, so that the part of the new echo signal containing the structural information is spliced ​​together; the mask matrix generation module generates a corresponding mask matrix according to the mask information of the input echo signal to be measured. code matrix; the improved Transformer encoder is combined with the mask matrix to extract features of the input echo signal to be tested, and the Transformer encoder output corresponding to the echo containing the target structure information is used as the input of the recognition module to finally obtain the target recognition result; in this way, the mask information of the missing part of the radar signal echo is used to assist the network to focus on the part of the echo information containing the target structure, and ignore the part of the echo information not containing the target structure. According to the mask information, the missing echo value is prevented from interfering with the target recognition process, thereby improving the recognition performance of the model for missing signals. When the radar echo signal is missing, it still has a good recognition effect and has good recognition robustness.

[0020] In addition, compared with the prior art solution of multiple short-dwell observations of a target, this embodiment is not limited to the time of observing the target and has a wider range of application scenarios.

[0021] In an optional embodiment of the present invention, in the mask information of the echo signal to be measured, the portion containing the target structure information is encoded as 1, and the portion not containing the target structure information is encoded as 0.

[0022] Specifically, in this embodiment, the echo signal obtained from the radar system, because it is necessary to take into account the detection of targets in the entire airspace, the echo signal for observing a single target may be discontinuous in time, and there may be missing values ​​in some places of the echo. Since the missing values ​​in the input data do not contain target structure information, inputting them into the model will affect the recognition performance of the model. Therefore, the part containing the target structure information is encoded as 1, and the part not containing the target structure information is encoded as 0.

[0023] In an optional embodiment of the present invention, please continue to refer to Figure 2As shown in the figure, the improved Transformer encoder includes a position encoding module, an embedding layer module, and a multi-layer improved Transformer encoder module connected in sequence; the position encoding module includes a sine-cosine position encoding, which is used to add position information to the data by splicing the generated position encoding; the embedding layer module is used to obtain a high-dimensional representation of each sample, and the number of neurons is set to 9×128;

[0024] The improved Transformer encoder module consists of a masked multi-head self-attention layer and a feedforward network layer connected sequentially.

[0025] It should be noted that Figure 2 The illustrated embodiment only schematically illustrates the configuration of three improved Transformer encoder modules.

[0026] In an optional embodiment of the present invention, please continue to refer to Figure 2 As shown, the feedforward network layer includes a first fully connected layer, a ReLU activation function, and a second fully connected layer; wherein, the number of neurons in the first fully connected layer is set to 128×256, and the number of neurons in the second fully connected layer is set to 256×128.

[0027] In an optional embodiment of the present invention, the position encoding uses sine-cosine position encoding, whose output dimension is 8, and the generated position encoding is added to the data by splicing; the masked multi-head self-attention layer uses the mask information. The length of the QKV vector in the self-attention layer is set to 100, and the number of self-attentions of the multi-head self-attention is 6. The calculation formula is:

[0028]

[0029] Where M∈R D×D , D is the number of samples contained in the input data, when the i,jth sample 1≤i,j≤D, x i and x j When both are not missing, M[i,j]=1, when x i and x j When at least one is missing, M[i,j]=0, ​​let the input matrix X=[x1,x2,…x i ,…x m ], where the i-th input vector x i ∈R D×1 , m is the number of input vectors, and then three parameter matrices are generated respectively and Project the input matrix into three different spaces to obtain the query matrix Q = W q X=[q1,q2,…,qi ,…q m ], where the i-th query vector is Key matrix K = W k X=[k1,k2,…,k i ,…k m ], where the i-th key vector is Value matrix V = W v X=[v1,v2,…,v i ,…v m ], where the i-th value vector is D k is the matrix W q , matrix W k and matrix W v The number of matrix rows.

[0030] In an optional embodiment of the present invention, please continue to refer to Figure 2 As shown, the recognition module includes a fully connected layer and a SoftMax activation function connected in sequence;

[0031] The number of neurons in the fully connected layer is set to 128×X, where X is the number of recognition categories.

[0032] In an optional embodiment of the present invention, the training process of the mask information auxiliary network includes:

[0033] Obtain an original training data set; wherein the original training data set includes multiple original training samples, and each original training sample is marked with a corresponding category label;

[0034] Obtain the network training data set and the mask information corresponding to the network training data set;

[0035] The mask information auxiliary network is trained using the network training data set, the mask information corresponding to the network training data set, and the category labels corresponding to the original training samples to obtain a trained mask information auxiliary network.

[0036] In an optional embodiment of the present invention, the process of obtaining the original training sample set includes:

[0037] Acquire X categories of echo signals, where each category of echo signals includes Z echo signals; where X ≥ 3 and Z ≥ 10000;

[0038] Performing modulo-2 norm normalization processing on the amplitude of each echo signal to obtain a normalized echo signal;

[0039] Taking the modulus value of each normalized echo signal to obtain the original training sample;

[0040] The modulus values ​​of all normalized echo signals are taken to obtain the original training data set.

[0041] In an optional embodiment of the present invention, the process of obtaining the network training data set and the mask information corresponding to the network training data set includes:

[0042] Obtain the missing rate μ of the original training sample position in the original training data set, randomly set μ×L positions of each original training sample to zero, and record the current zero position; where μ is expressed as:

[0043] μ=D lack / D∈[0,1);

[0044] Among them, D lack is the number of original training samples with missing positions, and D is the total number of original training samples in the original training dataset;

[0045] Add the corresponding category label to each randomly set zero original training sample and record the mask information corresponding to the zero position;

[0046] Add corresponding category labels to all randomly set zero original training samples, and record the mask information corresponding to the zero position to obtain the network training data set and the mask information corresponding to the network training data set.

[0047] In an optional embodiment of the present invention, the mask information auxiliary network is trained using the network training data set, the mask information corresponding to the network training data set, and the category labels corresponding to the original training samples, to obtain the trained mask information auxiliary network, including:

[0048] The mask information corresponding to the network training dataset is input into the mask matrix generation module to generate a mask matrix; wherein the mask matrix acts on the masked multi-head self-attention layer in the improved Transformer encoder module;

[0049] Input the network training data set into the improved Transformer encoder module to obtain the processed echo signal;

[0050] The processed echo signal is input into the recognition module to obtain the predicted category label;

[0051] The cross entropy loss function is used to calculate the loss between the predicted category label and the category label of the original training sample, and then the back propagation algorithm is used to iteratively update the network parameters until the cross entropy loss function converges to obtain a trained mask information auxiliary network.

[0052] In an optional embodiment of the present invention, the expression of the cross entropy loss function is:

[0053]

[0054] Where H() is the cross entropy loss function, Q is the predicted category label of the multi-frequency mixed training network assisted by repetition information, P is the true category label of the training sample in the training dataset, x is the category number of the training sample in the network training dataset, x = 1, 2, ..., X, X is the total number of categories of training samples in the network training dataset, log is the logarithm operation with base 10, and N is the batch size in one epoch.

[0055] It should be noted that, in the mask-assisted information auxiliary network of the present embodiment, the mask information of the echo signal is input as auxiliary information into the mask matrix generation module, and the mask matrix output by the mask-assisted information auxiliary network is applied to the multi-head self-attention part in the Transformer encoder group module. During the training process of the mask-assisted information auxiliary network of the present embodiment, for missing signals, the missing values ​​do not contain target information and are not helpful for the target recognition task. Inputting them into the model will affect the model performance. In order to avoid the interference of missing values ​​on the recognition results, in the multi-head self-attention layer, the feature corresponding to the observation value should be zero in the attention distribution of the feature corresponding to the missing value. When calculating the self-attention output in the multi-head self-attention layer, it is necessary to perform a mask operation on the attention distribution. The calculation formula is as follows:

[0056]

[0057]

[0058] Among them, M is the mask matrix, L is the number of elements in the input data, and D k is the matrix W q , matrix W k and matrix W v The number of matrix rows.

[0059] In this embodiment, the mask information of the echo signal to be measured and the echo signal to be measured are respectively input into the Transformer encoder and recognition module of the trained mask information auxiliary network, and the probability of the echo signal to be measured being identified as each category is calculated through the SoftMax layer, and the category corresponding to the highest probability is selected as the recognition result.

[0060] In an optional embodiment of the present invention, a simulation experiment is conducted to demonstrate the recognition effect of a mask-assisted network robust target recognition method under time resource constraints proposed by the present invention.

[0061] 1. Experimental conditions:

[0062] The hardware platform of the simulation experiment of this embodiment is: the processor is Intel(R) Core(TM) i7-10700 CPU@2.90GHZ and the memory is 32GB.

[0063] The software platform for the simulation experiment of this embodiment is: Windows 10 operating system and Python 3.7.

[0064] The data used in the simulation experiment of this embodiment are 12 models of three types of aircraft generated by the electromagnetic simulation software CST, namely helicopters, propellers and jets, with 4 models of each of the three types of aircraft. The specific rotor physical parameters are shown in Table 1.

[0065] Table 1 Aircraft rotor physical parameters

[0066]

[0067]

[0068] The original training dataset radar echo signals used in the simulation experiments of this embodiment were generated at an operating frequency of 8 GHz, a dwell time of 100 ms, and a pulse repetition frequency of 6 kHz. Each aircraft model contained 30,000 radar echo signals, and each aircraft category contained 120,000 radar echo signals, for a total of 360,000 radar echo signals. In the experiment, Gaussian white noise with a signal-to-noise ratio of 0 dB was added to the original training dataset, where the signal-to-noise ratio is defined as the ratio of the micro-motion component to the noise.

[0069] The radar echo signals of the network training data set used in the simulation experiment of this embodiment are generated from the radar echo signals of the original training data set; the random missing rate μ is set to 0.2, and the length of the radar echo signals in each original training data set is 600, that is, 120 positions are randomly selected between the radar echo signals of the original training data set [1,600] as missing values ​​without target structure information, and the 120 randomly selected position records are used to generate mask information and the corresponding parts of the radar echo signals of the original training data set are set to zero according to the positions, and the corresponding category information is added.

[0070] The simulation experiment of this embodiment compares the recognition results of the mask information auxiliary network, the unmasked information auxiliary network and the recurrent neural network under different missing conditions; among them, the unmasked information auxiliary network has a similar structure to the mask information auxiliary network, and the mask information is not used in the multi-head attention layer; the input dimension of the recurrent neural network is 10, the hidden layer dimension is set to 128, and the number of layers is 1.

[0071] The radar echo signal test data set used in the simulation experiment of this embodiment was generated at a radar operating frequency of 8 GHz, a pulse repetition frequency of 6 kHz, and a dwell time of 100 ms. The test data sets were divided into four different groups based on the echo loss rate. The first group was set to have no echo information loss, with a loss rate of zero; the second group was set to have a maximum of 10 ms of loss after every 10 ms of target observation, with a loss rate of 50%; the third group was set to have a maximum of 20 ms of loss after every 10 ms of target observation, with a loss rate of 60%; and the fourth group was set to have a maximum of 40 ms of loss after every 10 ms of target observation, with a loss rate of 80%. Each sub-test data set contained 3,000 radar echo signals, meaning the test data set contained a total of 12,000 radar echo signals. In the experiment, Gaussian white noise with a signal-to-noise ratio of 0 dB was added to the training data set, where the signal-to-noise ratio is defined as the signal-to-noise ratio of the micro-motion component to the noise.

[0072] 2. Experimental Content and Results

[0073] See Figure 3 As shown, Figure 3 The following is a simulation experiment result graph provided by an embodiment of the present invention, where the green line represents the recognition results of the unmasked information-assisted network under different missing data conditions, the red line represents the recognition results of the RNN network under different missing data conditions, and the blue line represents the recognition results of the masked information-assisted network under different missing data conditions. It can be seen that the classification and recognition performance of the two Transformer network-based classification models is relatively close when the test data missing rate is 0 and 50%. However, as the test data missing rate increases, the recognition accuracy of the RNN network and the unmasked information-assisted network decreases significantly, while the recognition results of the masked information-assisted network are more robust, with a relatively smaller decline. At a missing data rate of 80%, the average classification and recognition rates of the RNN network and the unmasked information-assisted network decrease significantly compared to the masked information-assisted network. Overall, in scenarios where the test missing rate is inconsistent with the training rate, the masked information-assisted network performs better than the other two classification and recognition models under different missing data conditions. When the test missing rate differs significantly from the training rate, the recognition performance decreases relatively slowly, indicating that the masked information-assisted network-based recognition model has a certain degree of recognition robustness.

[0074] It should be noted that, in this document, relational terms such as first and second are used solely to distinguish one entity or operation from another, and do not necessarily require or imply any actual relationship or order between these entities or operations. Furthermore, the terms "comprise," "include," or any other variations thereof are intended to encompass non-exclusive inclusion, such that an article or device comprising a list of elements includes not only those elements but also other elements not explicitly listed. Without further limitation, an element defined by the phrase "comprising a..." does not preclude the presence of additional identical elements in the article or device comprising the element. Terms such as "connected" or "connected" are not limited to physical or mechanical connections but may include electrical connections, whether direct or indirect. References to orientations or positional relationships, such as "upper," "lower," "left," and "right," are based on the orientations or positional relationships shown in the accompanying drawings and are intended solely to facilitate description and simplify the description of the present invention. They do not indicate or imply that the device or element referred to must have, be constructed, or operate in a specific orientation, and are therefore not to be construed as limiting the present invention.

[0075] 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 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.

[0076] 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 mask-assisted network robust radar target recognition method under time resource constraints, characterized in that: include: Acquire an echo signal to be measured; wherein the echo signal to be measured includes mask information, and in the mask information, a portion containing target structure information is encoded as 1, and a portion not containing target structure information is encoded as 0; The echo signal to be measured and the mask information of the echo signal to be measured are input into the trained mask information auxiliary network to obtain a recognition result; wherein, the mask information auxiliary network includes an improved Transformer encoder, an identification module and a mask matrix generation module, the mask matrix generation module generates a mask matrix according to the mask information of the echo signal to be measured, the improved Transformer encoder combines the mask matrix to extract features of the echo signal to be measured, and outputs the echo signal after feature extraction; the identification module identifies the echo signal after feature extraction and outputs a recognition result; wherein, The improved Transformer encoder includes a position encoding module, an embedding layer module and a multi-layer improved Transformer encoder module connected in sequence; wherein, the position encoding module includes a sine-cosine position encoding, which is used to add position information to the data by splicing the generated position encoding; the embedding layer module is used to obtain a high-dimensional representation of each sample; the improved Transformer encoder module includes a masked multi-head self-attention layer and a feedforward network layer connected in sequence.

2. The mask-assisted network robust radar target recognition method under time resource constraints according to claim 1 is characterized in that The feedforward network layer includes a first fully connected layer, a ReLU activation function, and a second fully connected layer; wherein the number of neurons in the first fully connected layer is set to , the number of neurons in the second fully connected layer is set to .

3. The mask-assisted network robust radar target recognition method under time resource constraints according to claim 1 is characterized in that The recognition module includes a fully connected layer and a SoftMax activation function connected in sequence; Among them, the number of neurons in the fully connected layer is set to , is the number of identified categories.

4. The mask-assisted network robust radar target recognition method under time resource constraints according to claim 1 is characterized in that The training process of the mask information auxiliary network includes: Obtaining an original training data set; wherein the original training data set includes a plurality of original training samples, each of the original training samples is marked with a corresponding category label; Obtaining the network training data set and mask information corresponding to the network training data set; The mask information auxiliary network is trained using the network training data set, the mask information corresponding to the network training data set, and the category labels corresponding to the original training samples to obtain a trained mask information auxiliary network.

5. The mask-assisted network robust radar target recognition method under time resource constraints according to claim 4 is characterized in that: The process of obtaining the original training sample set includes: Get The echo signals of each category include echo signals; among them, , ; Performing modulo-2 norm normalization processing on the amplitude of each echo signal to obtain a normalized echo signal; Taking a modulus value of each normalized echo signal to obtain an original training sample; The modulus values ​​of all the normalized echo signals are taken to obtain the original training data set.

6. The mask-assisted network robust radar target recognition method under time resource constraints according to claim 5 is characterized in that: The process of acquiring the network training data set and the mask information corresponding to the network training data set includes: Obtain the missing rate of the original training sample position in the original training data set , for each of the original training samples Randomly set the positions to zero and record the current zero position; among them, The expression is: ； in, is the number of original training samples with missing positions, is the total number of original training samples in the original training data set; Adding a corresponding category label to each randomly set zeroed original training sample, and recording mask information corresponding to the zeroed position; Add corresponding category labels to all the original training samples that are randomly set to zero, and record mask information corresponding to the zero-setting positions to obtain the network training data set and the mask information corresponding to the network training data set.

7. The mask-assisted network robust radar target recognition method under time resource constraints according to claim 6 is characterized in that: The step of training the mask information auxiliary network using the network training data set, the mask information corresponding to the network training data set, and the category labels corresponding to the original training samples to obtain a trained mask information auxiliary network includes: Inputting the mask information corresponding to the network training data set into the mask matrix generation module to generate a mask matrix; wherein the mask matrix acts on the masked multi-head self-attention layer in the improved Transformer encoder module; Inputting the network training data set into the improved Transformer encoder module to obtain a processed echo signal; Inputting the processed echo signal into a recognition module to obtain a predicted category label; A cross-entropy loss function is used to calculate the loss between the predicted category label and the category label of the original training sample, and then a back-propagation algorithm is used to iteratively update the network parameters until the cross-entropy loss function converges, thereby obtaining the trained mask information auxiliary network.

8. The mask-assisted network robust radar target recognition method under time resource constraints according to claim 7 is characterized in that: The expression of the cross entropy loss function is: ； in, is the cross entropy loss function, The predicted category labels of the multi-frequency mixed training network assisted by re-frequency information, is the true category label of the training samples in the training dataset, is the category number of the training sample in the network training data set, , is the total number of categories of training samples in the network training dataset, For one in .

Images