Intelligent target identification method based on electromagnetic scattering characteristic multi-source information fusion

Abstract

The invention discloses a target intelligent identification method based on electromagnetic scattering characteristic multi-source information fusion, and the method comprises the steps: obtaining high-resolution range profile data and radar cross section data of a target, and carrying out the preprocessing; aligning the preprocessed high-resolution range profile data and radar cross section data according to a preset rule to obtain electromagnetic scattering data; according to the method, the high-resolution range profile data and the radar cross section data of the target are obtained respectively, the two kinds of data are aligned and then input into the target intelligent recognition model to obtain the recognition result of the target, richer feature information is provided for space target recognition, and the recognition accuracy of the target is improved. The method has a certain promotion effect on improvement of the accuracy rate of target individual recognition; in a real environment, under the condition that high-resolution range profile data or radar cross section data may be missing, missing data is randomly generated to serve as input of a target intelligent recognition model, and generalization of target intelligent recognition is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of space target recognition, and particularly relates to an intelligent target recognition method based on multi-source information fusion of electromagnetic scattering characteristics. Background Art

[0002] Electromagnetic scattering characteristic data has the characteristics of long distance, non-invasive, high information volume, etc., can fully reflect the characteristics of the target's scattering cross-section, scattering intensity, etc., and has the ability to support target recognition. Establishing a target recognition method based on electromagnetic scattering characteristics will help to achieve target recognition under all-weather conditions and improve the situation awareness ability. However, due to factors such as terrain, atmosphere, and stray signals, the distortion or loss of the target's scattering characteristics occurs, resulting in problems such as poor target recognition effect and insufficient utilization of data.

[0003] Existing recognition methods for intelligent recognition of targets based on electromagnetic scattering characteristic data mostly target a single data source. For example, in the existing patent 202210057748.X, a radar target recognition method based on ultra-wideband electromagnetic scattering characteristics effectively recognizes radar targets by using high-resolution range profiles (HRRP). In the existing patent 202211601940.7, a deep network target recognition method based on radar scenes effectively recognizes targets through radar cross-section (RCS). The data from a single source may not fully characterize the target. At the same time, there is also a phenomenon that some data of the target cannot be collected during the data acquisition process. Summary of the Invention

[0004] The purpose of the present invention is to propose an intelligent target recognition method based on multi-source information fusion of electromagnetic scattering characteristics and its manufacturing method to solve the problems raised in the background art.

[0005] To achieve the above purpose, the technical solutions adopted by the present invention are as follows:

[0006] An intelligent target recognition method based on multi-source information fusion of electromagnetic scattering characteristics and its manufacturing method proposed by the present invention includes:

[0007] Obtain the high-resolution range profile data and radar cross-section data of the target, and perform preprocessing;

[0008] Align the preprocessed high-resolution range profile data and radar cross-section data according to preset rules to obtain electromagnetic scattering data;

[0009] Input the electromagnetic scattering data into the target intelligent recognition model. In the target intelligent recognition model, the high-resolution range profile data in the electromagnetic scattering data passes through the first branch module to obtain the first feature of the high-resolution range profile data. The first branch module includes a preset number of OneDimConv modules, a Flatten layer, and a first MLP fully connected network connected in sequence from data input to output;

[0010] The radar cross-section data in the electromagnetic scattering data passes through the second branch module to obtain the second feature of the radar cross-section data, and the second branch module is a second MLP fully connected network;

[0011] The first feature and the second feature are fused, and then passed through a third MLP fully connected network to obtain the recognition result of the target.

[0012] Preferably, the obtaining of the high-resolution range profile data and radar cross-section data of the target and the preprocessing thereof include:

[0013] The obtained high-resolution range profile HRRP data is an HRRP data sequence: {H1, H2, H3, … H i , …, H n}, and the radar cross-section RCS data is an RCS data sequence: {R1, R2, R3, … R l , …, R k}, where n represents the number of frames in the HRRP data sequence, k represents the number of frames in the RCS data sequence, H i represents the i-th frame data in the HRRP data sequence, and R l represents the l-th frame data in the RCS data sequence;

[0014] Remove the pixel points with pixel values of zero in each frame of the HRRP data sequence;

[0015] Normalize the HRRP data sequence and RCS data sequence after removal. The normalized HRRP data sequence is The normalized RCS data sequence is

[0016] Preferably, the aligning of the preprocessed high-resolution range profile data and radar cross-section data according to a preset rule to obtain electromagnetic scattering data includes:

[0017] Each frame data in the normalized HRRP data sequence and each frame data in the RCS data sequence carry the time of obtaining each frame data and the height of the target distance receiving device at that time, where the receiving device is the device for obtaining the high-resolution range profile data and radar cross-section data of the target;

[0018] When aligning, for each frame data in the normalized HRRP data sequence, using the time and height corresponding to the frame data as a reference, filter the frame data with corresponding time and height from the normalized RCS data sequence within a preset range, and use this frame data and each frame data filtered from the normalized RCS data sequence together as a frame of electromagnetic scattering data, then an electromagnetic scattering data sequence is obtained: {D1, D2, D3, … D j , …, D n}, where the number of frames in the electromagnetic scattering data sequence is the same as that in the HRRP data sequence, and D j represents the j-th frame data in the electromagnetic scattering data sequence.

[0019] Preferably, the OneDimConv module includes a convolutional layer, an attention mechanism module, a batch normalization layer, and an activation layer connected in sequence, and the input of the OneDimConv module is fused with the input of the activation layer to obtain the output of the OneDimConv module;

[0020] In the OneDimConv module, the calculation formula corresponding to the convolutional layer is as follows:

[0021]

[0022] where L out represents the output of the convolutional layer in the OneDimConv module, padding represents the number of zeros padded on both sides of the input data of the convolutional layer in the OneDimConv module, kernel represents the size of the convolutional kernel in the convolutional layer, and stride represents the step size of the convolutional kernel movement in the convolutional layer.

[0023] Preferably, in the attention mechanism module, the input of the attention mechanism module passes through a global average pooling layer and a fully connected layer in sequence to obtain a third feature, and then the third feature is fused with the input of the attention mechanism module to obtain the output of the attention mechanism module.

[0024] Preferably, in the target intelligent recognition model, for each frame D j data in the electromagnetic scattering data sequence, the HRRP frame data in the D j data is first dimensionally expanded and then input into the first branch module to obtain the first feature of the HRRP data. The RCS frame data in the D j data is input into the second branch module to obtain the second feature of the RCS data, and then the first feature and the second feature are fused and input into the third MLP fully connected network to obtain the recognition result of each D j data. From all D jSelect the recognition result with the highest proportion in the recognition results of the data as the final target recognition result.

[0025] Preferably, in the target intelligent recognition model, for each frame D in the electromagnetic scattering data sequence j data, D j For the HRRP frame data in the data, first perform dimensional expansion, and then input it into the first branch module to obtain the first feature of the HRRP data. D j The RCS frame data in the data is input into the second branch module to obtain the second feature of the RCS data, and then the first feature and the second feature are fused and input into the third MLP fully connected network to obtain the recognition results of each D j data, and select the recognition result with the highest proportion from all the D j recognition results of the data as the final target recognition result.

[0026] Preferably, when the high-resolution range profile data or radar cross-section data of the target is missing, randomly initialize and generate the missing high-resolution range profile data or radar cross-section data, and then directly input the generated high-resolution range profile data or radar cross-section data and the non-missing radar cross-section data or high-resolution range profile data into the target intelligent recognition model to obtain the recognition result of the target.

[0027] Compared with the prior art, the beneficial effects of the present invention are:

[0028] The target intelligent recognition method based on multi-source information fusion of electromagnetic scattering characteristics obtains the high-resolution range profile data and radar cross-section data of the target respectively, aligns the two types of data, and then inputs them into the target intelligent recognition model to obtain the recognition result of the target, providing richer feature information for space target recognition, and having a certain promoting effect on improving the accuracy of target individual recognition;

[0029] The target intelligent recognition method based on multi-source information fusion of electromagnetic scattering characteristics aims at the situation where high-resolution range profile data or radar cross-section data may be missing in the real environment, randomly generates the missing data as the input of the target intelligent recognition model, and improves the generalization of the target intelligent recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 is a flowchart of the target intelligent recognition method based on multi-source information fusion of electromagnetic scattering characteristics of the present invention;

[0031] Figure 2 is a schematic structural diagram of the target intelligent recognition model of the present invention;

[0032] Figure 3 is a schematic structural diagram of the first branch module of the present invention;

[0033] Figure 4 It is a schematic structural diagram of the second branch module of the present invention;

[0034] Figure 5 It is a schematic structural diagram during the training of the first branch module of the present invention;

[0035] Figure 6 It is a schematic structural diagram during the training of the second branch module of the present invention. Specific embodiments

[0036] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0037] In one embodiment, as Figures 1-4 shown, a target intelligent recognition method based on multi-source information fusion of electromagnetic scattering characteristics is provided, including:

[0038] S1. Obtain the high-resolution range profile data and radar cross-section data of the target, and perform preprocessing (specifically, perform removal processing on the HRRP data, and perform normalization processing on the HRRP data and RCS data);

[0039] It should be noted that in this embodiment, a detection device (such as a broadband radar) is used to obtain the high-resolution range profile data and radar cross-section data of the target to be measured within a period of time; among them, the obtained high-resolution range profile HRRP data is the HRRP data sequence: {H1, H2, H3,... H i ,..., H n}, and the radar cross-section RCS data is the RCS data sequence: {R1, R2, R3,... R l ,..., R k}, where n represents the number of frames in the HRRP data sequence, k represents the number of frames in the RCS data sequence, H i represents the i-th frame data in the HRRP data sequence, and R l represents the l-th frame data in the RCS data sequence; the lengths of the HRRP data sequence and the RCS data sequence can be adjusted according to actual needs; among them, the high-resolution range profile data is the HRRP data, and the radar cross-section data is the RCS data;

[0040] Remove the pixel points with pixel value of zero in each frame data of the HRRP data sequence; before removal, there are 4096 pixel points in each HRRP frame data (where the HRRP frame data is a frame of data in the HRRP data sequence). After removing the invalid information on both sides in each HRRP frame data (i.e., removing the pixel points with pixel value of zero), there are 256 pixel points in each HRRP frame data, and there is only one pixel point in each RCS frame data (where the RCS frame data is a frame of data in the RCS data sequence).

[0041] Normalize the removed HRRP data sequence and RCS data sequence, and the normalized HRRP data sequence is The normalized RCS data sequence is In this implementation, the maximum-minimum normalization operation is adopted to eliminate the influence of dimension.

[0042] S2. Align the preprocessed high-resolution range profile data and radar cross-section data according to preset rules to obtain electromagnetic scattering data, specifically including:

[0043] Each frame data in the normalized HRRP data sequence and each frame data in the RCS data sequence carry the time of obtaining each frame data and the height of the target distance receiving device at that time, where the receiving device is the device for obtaining the high-resolution range profile data and radar cross-section data of the target;

[0044] When aligning, for each frame data in the normalized HRRP data sequence, with the time and height corresponding to the frame data as the benchmark, screen the frame data with the corresponding time and height from the normalized RCS data sequence within a preset range, and use this frame data and the frame data screened from the normalized RCS data sequence together as a frame of electromagnetic scattering data, then an electromagnetic scattering data sequence is obtained: {D1, D2, D3, … D j , …, D n}, where the number of frames in the electromagnetic scattering data sequence is the same as the number of frames in the HRRP data sequence, and D j represents the j-th frame data in the electromagnetic scattering data sequence.

[0045] It should be noted that, such as The time corresponding to the frame data is the 60th second, and the height is 300 kilometers. The preset range of time is plus or minus 0.05 seconds, and the preset range of height is plus or minus 0.5 kilometers. Then, the frame data within the time range of (60 - 0.05) to (60 + 0.05) seconds and the height range of (300 - 0.5) to (300 + 0.5) meters are screened out from the normalized RCS data sequence (that is, the frame data within the time range of 59.95 to 60.05 seconds and the height range of 299.5 to 300.5 meters are screened out from the normalized RCS data sequence). Then, the frame data screened out from the normalized RCS data sequence is combined with The frame data is used as a frame of electromagnetic scattering data D1. The alignment process of each frame of data is the same and will not be described repeatedly. Moreover, there are duplicate normalized frame RCS data in each frame of the aligned electromagnetic scattering data.

[0046] S3. Input the electromagnetic scattering data into the target intelligent recognition model. In the target intelligent recognition model, the high-resolution range profile data in the electromagnetic scattering data passes through the first branch module to obtain the first feature of the high-resolution range profile data. The first branch module includes a preset number of OneDimConv modules, a Flatten layer, and a first MLP fully connected network connected in sequence from data input to output (this step is used to extract features from the high-resolution range profile data in the electromagnetic scattering data);

[0047] Among them, the OneDimConv module includes a convolutional layer (using a one-dimensional convolutional layer), an attention mechanism module, a batch normalization layer, and an activation layer (using the Relu activation function) connected in sequence. Moreover, the input of the OneDimConv module is fused with the input of the activation layer (this fusion is an addition operation) to obtain the output of the OneDimConv module;

[0048] The calculation formula corresponding to the convolutional layer in the OneDimConv module is as follows:

[0049]

[0050] Among them, L out represents the output of the convolutional layer in the OneDimConv module, padding represents the number of zeros padded on both sides of the input data of the convolutional layer in the OneDimConv module (padding zeros makes the scale of the feature vector unchanged before and after convolution), kernel represents the size of the convolutional kernel in the convolutional layer, and stride represents the step size of the convolutional kernel moving in the convolutional layer;

[0051] In the attention mechanism module, the input of the attention mechanism module sequentially passes through a global average pooling layer and a fully connected layer to obtain a third feature, and then the third feature is fused with the input of the attention mechanism module (this fusion is a multiplication operation) to obtain the output of the attention mechanism module;

[0052] It should be noted that in the target intelligent recognition model, for each frame D in the electromagnetic scattering data sequence j data, the HRRP frame data in D j data is first dimensionally expanded (such as using the Unsqueeze operation to expand the dimension), and then input into the first branch module to obtain HRRP extraction features (i.e., the first feature);

[0053] The number of OneDimConv modules in the first branch module is N, and the number of layers of the fully connected layer in the first MLP fully connected network is K. The feature extraction of HRRP is realized through the first branch module, and the HRRP extraction feature is expressed as The specific formula is as follows:

[0054]

[0055] In this embodiment, N is 4 and K is 2. Before inputting the HRRP frame data in D j data into the HRRP branch module, the HRRP frame data in D j data is first dimensionally expanded. For example, the dimension of the HRRP frame data in D j data is [1, 256], and after dimensional expansion, it is [1, 256, 1]; the convolutional layer in the first OneDimConv module uses a convolutional kernel of size 1*64, the stride is 1, and the number of output convolutional kernels is 32, realizing the conversion of data with a size of [1, 256, 1] into a feature vector of [1, 256, 32]; the convolutional layer in the second OneDimConv module uses a convolutional kernel of size 1*32, the stride is 1, and the number of output convolutional kernels is 32. The convolutional layers in the third OneDimConv module and the fourth OneDimConv module use convolutional kernels of size 1*16 and a stride of 2, and the number of output convolutional kernels is 32. After passing through four OneDimConv modules, a feature vector of size [1, 64, 32] is obtained, and then through the Flatten layer, the feature vector of [1, 64, 32] is converted into a two-dimensional matrix of [1, 2048]. Finally, through the first MLP fully connected network, a feature of [1, 256] dimension is obtained, which is the output of the first branch module (i.e., the HRRP extraction feature).

[0056] S4. The radar cross-section data in the electromagnetic scattering data obtains the second feature of the radar cross-section data through the second branch module, and the second branch module is the second MLP fully connected network (this step is used for feature extraction of the radar cross-section data in the electromagnetic scattering data);

[0057] Among them, for each frame D in the electromagnetic scattering data sequence j data, the RCS frame data in D j data is input into the second branch module to obtain the RCS extraction feature (i.e., the second feature);

[0058] In this embodiment, the number of fully connected layers in the second MLP fully connected network is 2. For example, if the dimension of the RCS frame data in the D j data input into the second branch module is [1, 128], after passing through the two fully connected layers in the second MLP fully connected network, the dimensions are sequentially transformed into [1, 64] and [1, 32], and the feature of dimension [1, 32] is the output of the second branch module (i.e., the RCS extraction feature).

[0059] S5. The first feature and the second feature are fused, and then the recognition result of the target is obtained through the third MLP fully connected network;

[0060] Among them, for each frame D in the electromagnetic scattering data sequence j data, the HRRP extraction feature and the RCS extraction feature obtained from each frame D j data are fused (this fusion is a splicing operation) and input into the third MLP fully connected network to obtain the recognition result of each D j data. The recognition result with the highest proportion is selected from the recognition results of all D j data as the final target recognition result; in this embodiment, the number of fully connected layers in the third MLP fully connected network is 4, the dimension of the HRRP extraction feature is [1, 256], the dimension of the RCS extraction feature is [1, 32], the two are spliced to obtain a splicing feature with a dimension of [1, 288], and then through the third MLP fully connected network, a target recognition result with a dimension of [1, 21] is output.

[0061] In another embodiment, such as Figures 5-6As shown, during the training process of the target intelligent recognition model, the target intelligent recognition model further includes a fourth MLP fully connected network and a fifth MLP fully connected network. The output of the first branch module (i.e., the first feature) is input into the fourth MLP fully connected network to obtain the recognition result of the first branch module, and the recognition result of the first branch module is used to calculate the loss of the first branch module. The output of the second branch module (i.e., the second feature) is input into the fifth MLP fully connected network to obtain the recognition result of the second branch module, and the recognition result of the second branch module is used to calculate the loss of the second branch module;

[0062] The loss function of the target intelligent recognition model includes the loss of the first branch module, the loss of the second branch module, and the loss of the third MLP fully connected network. The specific formula of the loss function Loss of the target intelligent recognition model is as follows:

[0063] Loss = α × Loss RCS + β × Loss HRRP + γ × Loss concat ;

[0064] Among them, Loss RCS represents the loss of the first branch module, Loss HRRP represents the loss of the second branch module, Loss concat represents the loss of the third MLP fully connected network. α, β, and γ represent the weights of the loss of the second branch module, the loss of the first branch module, and the loss of the third MLP fully connected network in sequence. In this embodiment, α is taken as 1, β is taken as 1, and γ is taken as 3. In this embodiment, each loss function adopts the cross-entropy loss function.

[0065] When the high-resolution range image data or radar cross-section data of the target is missing, randomly initialize and generate the missing high-resolution range image data or radar cross-section data, and then directly input the generated high-resolution range image data or radar cross-section data and the radar cross-section data or high-resolution range image data without missing into the target intelligent recognition model to obtain the recognition result of the target;

[0066] Among them, when the HRRP data of the target is obtained by the receiving device and the RCS data is missing, the RCS data is randomly initialized and generated by Gaussian noise. Subsequently, the real HRRP data (i.e., the existing HRRP data itself) and the generated RCS data are respectively input into the corresponding branches, and then the results of each branch are spliced and input into the third MLP fully connected network to obtain the recognition result of the target;

[0067] When the RCS data of the target is obtained by the receiving device while the HRRP data is missing, the HRRP data is randomly initialized by Gaussian noise. Subsequently, the real RCS data (i.e., the existing RCS data itself) and the generated HRRP data are respectively input into the corresponding branches, and then the results of each branch are concatenated and input into the third MLP fully connected network to obtain the recognition result of the target.

[0068] It should be understood that although Figure 1 the steps in the flowchart of Figure 1 are shown in sequence according to the indication of the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise clearly stated in this article, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover,

[0069] The above-described embodiments only represent several implementation manners of the present application, and their descriptions are relatively specific and detailed, but should not be construed as limiting the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the appended claims.

Claims

1. An object intelligent recognition method based on multi-source information fusion of electromagnetic scattering characteristics, characterized in that: The target intelligent recognition method based on multi-source information fusion of electromagnetic scattering characteristics includes: Obtain the high-resolution range profile data and radar cross-section data of the target, and perform preprocessing; Align the preprocessed high-resolution range profile data and radar cross-section data according to preset rules to obtain electromagnetic scattering data; Input the electromagnetic scattering data into the target intelligent recognition model. In the target intelligent recognition model, the high-resolution range profile data in the electromagnetic scattering data passes through the first branch module to obtain the first feature of the high-resolution range profile data. The first branch module includes a preset number of OneDimConv modules, a Flatten layer, and a first MLP fully connected network connected in sequence from data input to output; The radar cross-section data in the electromagnetic scattering data passes through the second branch module to obtain the second feature of the radar cross-section data, and the second branch module is a second MLP fully connected network; Fuse the first feature and the second feature, and then pass through the third MLP fully connected network to obtain the recognition result of the target.

2. The target intelligent recognition method and its manufacturing method based on multi-source information fusion of electromagnetic scattering characteristics according to claim 1, characterized in that: The obtaining of the high-resolution range profile data and radar cross-section data of the target, and the preprocessing include: The acquired high-resolution range profile (HRRP) data is the HRRP data sequence: {H1, H2, H3, … H i , …, H n}, and the radar cross-section (RCS) data is the RCS data sequence: {R1, R2, R3, … R l , …, R k}, where n represents the number of frames in the HRRP data sequence, k represents the number of frames in the RCS data sequence, H i represents the i-th frame data in the HRRP data sequence, and R l represents the l-th frame data in the RCS data sequence; Remove the pixel points with pixel values of zero in each frame of data in the HRRP data sequence; Normalize the HRRP data sequence and the RCS data sequence after removal, and the normalized HRRP data sequence is The normalized RCS data sequence is 3. The target intelligent recognition method and its manufacturing method based on multi-source information fusion of electromagnetic scattering characteristics according to claim 2, characterized in that: The aligning of the preprocessed high-resolution range profile data and radar cross-section data according to preset rules to obtain electromagnetic scattering data includes: Each frame of data in the normalized HRRP data sequence and each frame of data in the RCS data sequence carry the time of obtaining each frame of data and the height of the target from the receiving device at that time. The receiving device is the device for obtaining the high-resolution range profile data and radar cross-section data of the target; When aligning, for each frame data in the normalized HRRP data sequence, with the time and height corresponding to the frame data as the reference, screen the frame data with the corresponding time and height from the normalized RCS data sequence within a preset range, and use this frame data and each frame data screened from the normalized RCS data sequence together as a frame of electromagnetic scattering data, then an electromagnetic scattering data sequence is obtained: {D1, D2, D3, … D j , …, D n}, where the number of frames in the electromagnetic scattering data sequence is the same as the number of frames in the HRRP data sequence, and D j represents the j-th frame data in the electromagnetic scattering data sequence.

4. The target intelligent recognition method based on multi-source information fusion of electromagnetic scattering characteristics and its manufacturing method according to claim 1, characterized in that: The OneDimConv module includes a convolutional layer, an attention mechanism module, a batch normalization layer, and an activation layer connected in sequence. The input of the OneDimConv module is fused with the input of the activation layer to obtain the output of the OneDimConv module; The calculation formula corresponding to the convolutional layer in the OneDimConv module is as follows: Among them, L out represents the output of the convolutional layer in the OneDimConv module, padding represents the number of zeros padded on both sides of the input data of the convolutional layer in the OneDimConv module, kernel represents the size of the convolutional kernel in the convolutional layer, and stride represents the step size of the convolutional kernel movement in the convolutional layer.

5. The target intelligent recognition method and its manufacturing method based on multi-source information fusion of electromagnetic scattering characteristics according to claim 4, characterized in that: In the attention mechanism module, the input of the attention mechanism module passes through a global average pooling layer and a fully connected layer in sequence to obtain a third feature, and then the third feature is fused with the input of the attention mechanism module to obtain the output of the attention mechanism module.

6. The target intelligent recognition method and its manufacturing method based on multi-source information fusion of electromagnetic scattering characteristics according to claim 3, characterized in that: In the target intelligent recognition model, for each frame D in the electromagnetic scattering data sequence j data, D j For the HRRP frame data in the data, it is first dimensionally expanded and then input into the first branch module to obtain the first feature of the HRRP data, D j The RCS frame data in the data is input into the second branch module to obtain the second feature of the RCS data, and then the first feature and the second feature are fused and input into the third MLP fully connected network to obtain each D j recognition result of the data, and the recognition result with the highest proportion is selected from all D j recognition results of the data as the final target recognition result.

7. The target intelligent recognition method based on multi-source information fusion of electromagnetic scattering characteristics and its manufacturing method according to claim 1, characterized in that: During the training process of the target intelligent recognition model, the target intelligent recognition model further includes a fourth MLP fully connected network and a fifth MLP fully connected network. The output of the first branch module is input into the fourth MLP fully connected network to obtain the recognition result of the first branch module, and the recognition result of the first branch module is used to calculate the loss of the first branch module. The output of the second branch module is input into the fifth MLP fully connected network to obtain the recognition result of the second branch module, and the recognition result of the second branch module is used to calculate the loss of the second branch module; The loss function of the target intelligent recognition model includes the loss of the first branch module, the loss of the second branch module, and the loss of the third MLP fully connected network. The specific formula of the loss function Loss of the target intelligent recognition model is as follows: Loss = α × Loss RCS + β × Loss HRRP + γ × Loos concat ; Among them, Loss RCS represents the loss of the first branch module, and Loss HRRP represents the loss of the second branch module, and Loss concat represents the loss of the third MLP fully connected network. α, β, and γ represent the weights of the loss of the second branch module, the loss of the first branch module, and the loss of the third MLP fully connected network, respectively.

8. The target intelligent recognition method based on multi-source information fusion of electromagnetic scattering characteristics and its manufacturing method according to claim 1, characterized in that: When the high-resolution range image data or radar cross-section data of the target is missing, randomly initialize and generate the missing high-resolution range image data or radar cross-section data, and then directly input the generated high-resolution range image data or radar cross-section data and the radar cross-section data or high-resolution range image data without missing into the target intelligent recognition model to obtain the recognition result of the target.

Images