Radar target multimode data fusion identification method based on knowledge embedding model

Abstract

The invention discloses a radar target multimode data fusion identification method based on a knowledge embedding model, and relates to the technical field of radars, and the method comprises the steps: obtaining to-be-identified radar echo data and corresponding domain knowledge data; inputting to-be-identified radar echo data and corresponding domain knowledge data into the trained radar target identification network for processing, extracting a range profile feature token in the to-be-identified radar echo data, and extracting a domain knowledge text feature token of the domain knowledge data; carrying out fusion identification on the range profile feature token and the domain knowledge text feature token to obtain a fusion identification result of the to-be-identified target; wherein the trained radar target recognition network is obtained by training the initial radar target recognition network by taking preset category data as a training data set. The target recognition performance can be improved.

Description

Technical Field

[0001] The invention belongs to the field of radar technology, and in particular relates to a radar target multi-mode data fusion recognition method based on a knowledge embedding model. Background Art

[0002] The radar one-dimensional high-resolution range profile (HRRP) describes the distribution of the target scattering center along the radar line of sight. It includes relatively detailed information such as the shape, structure, and radial size of the target, and can achieve high-resolution modeling of the target. Therefore, it is widely used in the field of radar target recognition. However, due to the increasing complexity of actual application scenarios, target recognition based on a single HRRP has problems such as insufficient utilization of external knowledge and cannot meet the needs of robust target recognition.

[0003] Patent CN202310427516.3 discloses a radar target recognition model training method, a radar target recognition method and a device. By utilizing a feature extraction network based on a series-parallel convolutional network and pooled attention, the efficiency of extracting effective information from the signal is improved. Patent CN202210201673.8 discloses a radar target fusion recognition method and system. By using a one-dimensional range image sub-network to extract one-dimensional range image features and a two-dimensional range image sub-network to extract two-dimensional range image features, and using a fusion network for fusion recognition, the accuracy of target recognition is improved. The above technology only uses radar echo data in the application, and the recognition method's ability to extract and utilize features of multiple information needs to be improved.

[0004] Therefore, there is an urgent need to provide a radar target recognition method to improve the above-mentioned defects. Summary of the invention

[0005] In order to solve the above problems existing in the prior art, the present invention provides a radar target multi-mode data fusion recognition method based on a knowledge embedding model. The technical problem to be solved by the present invention is achieved through the following technical solutions: In a first aspect, the present invention provides a radar target multi-mode data fusion recognition method based on a knowledge embedding model, comprising: Obtain radar echo data to be identified and corresponding domain knowledge data; Input the radar echo data to be identified and the corresponding domain knowledge data into the trained radar target recognition network for processing, extract the range image feature tokens in the radar echo data to be identified, extract the domain knowledge text feature tokens of the domain knowledge data, and fuse the range image feature tokens and the domain knowledge text feature tokens to obtain the fusion recognition result of the target to be identified; The trained radar target recognition network is obtained by training the initial radar target recognition network using preset category data as a training data set.

[0006] Beneficial effects of the present invention: The present invention provides a radar target multi-mode data fusion recognition method based on a knowledge embedding model. In the recognition process of radar targets, domain knowledge text data is introduced, so that the domain knowledge text data provides more target information complementary to the one-dimensional high-resolution range image. The domain knowledge text data assists the trained radar target recognition network to better recognize the one-dimensional high-resolution range image features, which helps to improve the radar target recognition performance.

[0007] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 It is a flow chart of a radar target multi-mode data fusion identification method based on a knowledge embedding model provided by an embodiment of the present invention; Figure 2 is a schematic diagram of a trained radar target recognition network provided by an embodiment of the present invention; Figure 3 is a schematic diagram of a trained first feature extraction module provided by an embodiment of the present invention; Figure 4 is a schematic diagram of a trained second feature extraction module provided by an embodiment of the present invention; Figure 5 It is a schematic diagram of a trained mutual attention feature fusion recognition module provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0009] The present invention is further described in detail below with reference to specific embodiments, but the embodiments of the present invention are not limited thereto.

[0010] See also Figure 1 , Figure 1 The present invention provides a flow chart of a radar target multi-mode data fusion identification method based on a knowledge embedding model. The present invention provides a radar target multi-mode data fusion identification method based on a knowledge embedding model, including: S101. Obtain radar echo data to be identified and corresponding domain knowledge data.

[0011] Specifically, in this embodiment, the radar echo data to be identified is a one-dimensional high-resolution range image, and the domain knowledge data is target measurement information, including the target's track information and radar cross section (RCS). The radar echo data to be identified and the corresponding domain knowledge data are three types of aircraft target data, and the three types of aircraft are composed of aircraft A, aircraft B, and aircraft C.

[0012] S102, inputting the radar echo data to be identified and the corresponding domain knowledge data into the trained radar target recognition network for processing, extracting the range image feature tokens in the radar echo data to be identified, extracting the domain knowledge text feature tokens from the domain knowledge data, and fusing the range image feature tokens and the domain knowledge text feature tokens to obtain the fusion recognition result of the target to be identified; The trained radar target recognition network is obtained by training the initial radar target recognition network using preset category data as a training data set.

[0013] Specifically, see Figure 2 , Figure 2 is a schematic diagram of a trained radar target recognition network provided by an embodiment of the present invention. In this embodiment, the trained radar target recognition network includes a trained first feature extraction module, a trained second feature extraction module and a trained mutual attention feature fusion recognition module; the radar echo data to be recognized and the corresponding domain knowledge data are input into the trained radar target recognition network for processing, the distance image feature token in the radar echo data to be recognized is extracted, the domain knowledge text feature token of the domain knowledge data is extracted, and the distance image feature token and the domain knowledge text feature token are fused and recognized to obtain a fusion recognition result of the target to be recognized, including: The trained first feature extraction module is used to extract features from the radar echo data to be identified, and a range image feature token is obtained; Using the trained second feature extraction module to extract features from the domain knowledge data, and obtaining domain knowledge text feature tokens; The trained mutual attention feature fusion recognition module is used to fuse and recognize the distance image feature tokens and domain knowledge text feature tokens to obtain the fused recognition result of the target to be recognized.

[0014] It should be noted that the token is a token.

[0015] For further information, see Figure 3 , Figure 3: is a schematic diagram of a trained first feature extraction module provided in an embodiment of the present invention. In this embodiment, the trained first feature extraction module includes a one-dimensional high-resolution range image preprocessing module and a range image feature extraction module; the trained first feature extraction module is used to extract features from radar echo data to be identified to obtain a range image feature token, including: A one-dimensional high-resolution range image preprocessing module is used to preprocess the radar echo data to be identified to obtain a processed one-dimensional range image; The range image feature extraction module is used to extract features from the processed one-dimensional range image to obtain a range image feature token.

[0016] Optionally, the one-dimensional high-resolution range image preprocessing module uses L2 norm for normalization and center alignment operations. The range image feature extraction module adopts a Transformer structure, which includes 1 one-dimensional convolution layer (convolution kernel size is 8, step size is 8), 1 layer normalization layer, 2 attention layers and 1 layer normalization layer in sequence; each attention layer includes 1 multi-head attention layer, 1 layer normalization layer, 1 feedforward layer, and 1 layer normalization layer in sequence; among them, the multi-head attention layer includes 3 linear layers (128 input nodes and 128 output nodes), and the feedforward layer includes 1 linear layer (128 input nodes and 512 output nodes), 1 GELU activation layer and 1 linear layer (512 input nodes and 128 output nodes) in sequence.

[0017] For further information, see Figure 4 , Figure 4 : is a schematic diagram of a trained second feature extraction module provided in an embodiment of the present invention. In this embodiment, the trained second feature extraction module includes a domain knowledge text preprocessing module and a domain knowledge text feature extraction module; the trained second feature extraction module is used to extract features from domain knowledge data to obtain domain knowledge text feature tokens, including: The domain knowledge text preprocessing module is used to preprocess the domain knowledge data to obtain the processed domain knowledge text token; The domain knowledge text feature extraction module is used to extract features from the processed domain knowledge text tokens to obtain domain knowledge text feature tokens.

[0018] Optionally, the domain knowledge text preprocessing module is used to describe the domain knowledge in text, for example, "The target attitude is 30 degree" is described as "The target attitude is 30 degree", and the text descriptions of other domain knowledge are similar. Then, a word segmenter is used to obtain a numerical vector corresponding to the text description. In this embodiment, the CLIP vocabulary and Byte PairEncoding word segmentation method are used to obtain domain knowledge text tokens with a dimension of 50×128, where 50 is the number of domain knowledge text tokens and 128 is the dimension of each token. The domain knowledge text feature extraction module adopts a Transformer structure, which includes 2 attention layers, 1 layer normalization layer and 1 feedforward layer in sequence, and each attention layer includes 1 multi-head attention layer. Among them, the multi-head attention layer includes 3 linear layers (128 input nodes and 128 output nodes); the feedforward layer includes 1 linear layer (128 input nodes and 512 output nodes), 1 GELU activation layer and 1 linear layer (512 input nodes and 128 output nodes) in sequence.

[0019] For further information, see Figure 5 , Figure 5 : is a schematic diagram of a trained mutual attention feature fusion recognition module provided by an embodiment of the present invention. In this embodiment, the trained mutual attention feature fusion recognition module includes multiple cascaded mutual attention fusion modules and fusion classification layers, wherein each mutual attention fusion module includes a first normalization layer, a second normalization layer, a mutual attention fusion layer, a third normalization layer, a fourth normalization layer, a first feedforward layer, and a second feedforward layer; the trained mutual attention feature fusion recognition module is used to perform feature fusion and recognition on the distance image feature token and the domain knowledge text feature token to obtain a fusion recognition result of the target to be recognized, including: For the first-level mutual attention fusion module, the first normalization layer is used to normalize the range image feature token to obtain the first feature; The first feature is processed using a mutual attention fusion layer to obtain the second feature; Add the second feature to the range image feature token to obtain the third feature; The third feature is normalized by the third normalization layer to obtain the fourth feature; The fourth feature is processed by the first feed-forward layer to obtain the fifth feature; Add the fifth feature to the third feature to obtain a one-dimensional high-resolution range image classification feature vector; The second normalization layer is used to normalize the domain knowledge text feature tokens to obtain the sixth feature; The sixth feature is processed using the mutual attention fusion layer to obtain the seventh feature; Add the seventh feature to the domain knowledge text feature token to obtain the eighth feature; The fourth normalization layer is used to normalize the eighth feature to obtain the ninth feature; The ninth feature is processed by the second feed-forward layer to obtain the tenth feature; Add the tenth feature to the eighth feature to obtain the domain knowledge text classification feature vector; The one-dimensional high-resolution range image classification feature vector is concatenated with the domain knowledge text classification feature vector to obtain the fusion feature; The fusion classification layer is used to classify the fused feature vector to obtain the fusion recognition result of the target to be identified.

[0020] In this embodiment, the training process of the trained radar target recognition network includes: Acquire data of multiple preset categories to construct a training data set; wherein the training data set includes multiple samples, each sample includes radar echo data and corresponding domain knowledge data; optionally, each data sample includes a radar one-dimensional high-resolution range image and corresponding domain knowledge data, and the dimension of the radar one-dimensional high-resolution range image is 1×256; the dimension of the domain knowledge data is 1×5, and the domain knowledge data includes target attitude, target pitch angle, target height, target speed, and target RCS; Input some samples in the training data set into the The radar target recognition network to be trained is trained to obtain the The prediction results of the classification output during the training process; According to The prediction results of the classification output during the training process are consistent with the The real label of the sample of the radar target recognition network to be trained is used to calculate the classification loss and serve as the The classification loss of the training process; at the same time, according to the The one-dimensional high-resolution range image classification feature vector output during the first training process is matched with the domain knowledge text classification feature vector to calculate the matching loss and use it as the first Matching loss of the training process; According to The classification loss and matching loss of the first training process are back-propagated to update the The network parameters of the radar target recognition network to be trained are obtained The radar target recognition network to be trained is iterated until the number of training times or the degree of convergence meets the preset conditions, and a trained radar target recognition network is obtained.

[0021] It should be noted that using The sum of the classification loss and matching loss of the training process is back-propagated to update the The network parameters of the radar target recognition network to be trained.

[0022] Optionally, the total number of training data set samples is 3000, each sample includes 1 radar one-dimensional high-resolution range image and 1 domain knowledge data, the domain knowledge data includes target attitude, target pitch angle, target altitude, target speed, target RCS. The training data set includes 3 types of targets, including 1000 samples of aircraft A, 1000 samples of aircraft B, and 1000 samples of aircraft C. The total number of test data set samples is 1500, and the composition of each sample is the same as the training data set, including 500 samples of aircraft A, 500 samples of aircraft B, and 500 samples of aircraft C.

[0023] Furthermore, in this embodiment, the mutual attention feature fusion recognition module in the initial radar target recognition network includes an initial mutual attention fusion module, an initial one-dimensional high-resolution range image classification layer, an initial domain knowledge text classification layer and an initial fusion classification layer; The process of obtaining the classification loss of the training process includes: The initial mutual attention fusion module is used to generate the One-dimensional high-resolution range image classification feature vector and domain knowledge text classification feature vector in the training process; The initial one-dimensional high-resolution range image classification layer is used to classify the The one-dimensional high-resolution range image classification feature vector generated during the first training process is classified to obtain the first prediction result, and the first prediction result is compared with the first prediction result of the training The real labels of the samples of the radar target recognition network to be trained are compared to obtain the first classification loss; The initial fusion classification layer is used to The fusion feature vector generated in the first training process is classified to obtain the second prediction result, and the second prediction result is compared with the first training process. The real labels of the samples of the radar target recognition network to be trained are compared to obtain the second classification loss; Use the initial domain knowledge text classification layer to The domain knowledge text classification feature vector generated in the first training process is classified to obtain the third prediction result, and the third prediction result is compared with the first training process. The real labels of the samples of the radar target recognition network to be trained are compared to obtain the third classification loss; Add the first classification loss, the second classification loss, and the third classification loss to get The classification loss of the training process.

[0024] Optionally, the initial mutual attention feature fusion recognition module includes an initial mutual attention fusion module, an initial one-dimensional high-resolution range image classification layer, an initial domain knowledge text classification layer and an initial fusion classification layer. Among them, the initial mutual attention fusion module includes two first-layer normalization layers, one mutual attention fusion layer, two second-layer normalization layers and two feedforward layers in sequence. The mutual attention fusion layer includes two linear layers (128 input nodes and 256 output nodes), two linear layers (128 input nodes and 512 output nodes), two linear layers (256 input nodes and 128 output nodes), two linear layers (128 input nodes and 512 output nodes), two GELU activation layers, and two linear layers (512 input nodes and 128 output nodes). The one-dimensional high-resolution range image classification layer includes 1 linear layer (128 input nodes and 128 output nodes), 1 GELU activation layer, and 1 linear layer (128 input nodes and 3 output nodes). The domain knowledge text classification layer includes 1 linear layer (128 input nodes and 128 output nodes), 1 GELU activation layer, and 1 linear layer (128 input nodes and 3 output nodes). The fusion classification layer includes 1 linear layer (256 input nodes and 128 output nodes), 1 GELU activation layer, and 1 linear layer (128 input nodes and 3 output nodes).

[0025] Furthermore, in this embodiment, the expression of the first classification loss, the second classification loss or the third classification loss is: ; in, means a first category loss, a second category loss or a third category loss, represents the true label, Represents the prediction result, Indicates different modes, They represent one-dimensional high-resolution range image modality, domain knowledge text modality and fusion modality respectively. represents the total number of categories, Represents the category index.

[0026] Furthermore, in this embodiment, the expression of matching loss is: ; in, represents the matching loss value, represents the one-dimensional high-resolution range image classification feature vector, Represents the domain knowledge text classification feature vector, represents transpose, Represents the L2 norm of a vector.

[0027] Optionally, the batch size during training is 128, the initial learning rate is 1e-4, the optimizer is AdamW, and the total number of iterations is set to 100.

[0028] In summary, the radar target recognition method based on language large model knowledge embedding fusion provided by the present invention has the following beneficial effects: First, compared with the radar target recognition method that only uses HRRP echoes, the radar target multi-mode data fusion recognition method based on the knowledge embedding model constructed by the present invention makes full use of domain knowledge, including target motion state knowledge and target RCS knowledge. Domain knowledge and echoes describe target characteristics from different dimensions, improving the accuracy of target modeling.

[0029] Second, the multi-mode data fusion identification method of radar targets based on the knowledge embedding model constructed by the present invention fully considers the characteristics of domain knowledge and HRRP echoes, and designs the optimal comprehensive utilization of information. For HRRP echoes, by using feature extractors that perceive local structures and global structures, the distribution characteristics of scattering centers are mined to construct HRRP echo feature token sequences. For domain knowledge, by constructing a textual knowledge representation strategy, the physical meaning of each dimensional knowledge is clarified, and the understanding and reasoning ability of the large language model is used to mine the deep-level knowledge meaning to construct a domain knowledge token sequence. Through targeted processing of HRRP echoes and domain knowledge, the ability to mine various information contained in HRRP echoes and domain knowledge is improved.

[0030] Third, the multi-mode data fusion and recognition method of radar targets based on the knowledge embedding model constructed by the present invention designs a HRRP-knowledge fusion module based on mutual attention. Compared with conventional fusion strategies, it enhances the interaction between HRRP echoes and domain knowledge, improves the effectiveness of fusion, and enhances the radar target recognition performance.

[0031] In an optional embodiment of the present invention, the effect of the radar target recognition method based on language large model knowledge embedding fusion provided in the above embodiment is verified by simulation experiments, specifically: 1. Simulation conditions The hardware platform of the simulation experiment of the present invention is: the processor is Intel (R) i7 CPU, the main frequency is 3.20GHz, the memory capacity is 64GB, the graphics card is Nvidia RTX 3090, and the video memory capacity is 24GB.

[0032] The software platforms for the simulation experiment of the present invention are: Python 3.8, Pytorch 2.0.

[0033] 2. Simulation content and results analysis In the simulation experiment of the present invention, the method proposed by the present invention and an existing technology are used to conduct target recognition experiments on three types of aircraft, and the recognition rates are compared. The existing technology refers to the paper "AnImage is Worth 16x16 Words: Transformers for Image Recognition at Scale" published by Alexey Dosovitskiy et al., which applies the Transformer network to target recognition. The original network is designed for two-dimensional image data. In this simulation experiment, its main structure is retained, and the two-dimensional operation is modified into a one-dimensional operation to adapt to one-dimensional high-resolution range image data. The recognition rate is used in the simulation experiment to evaluate the performance of the recognition model, which is defined as the ratio of the number of correctly recognized samples to the total number of test samples. The higher the recognition rate, the more correctly recognized samples there are, and the better the recognition performance of the model, as shown in Table 1.

[0034] Table 1 Comparison of target recognition accuracy between the proposed method and the prior art

[0035] By comparing the recognition rates of the present invention and the prior art, it can be seen that the present invention effectively improves the radar target recognition performance.

[0036] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that there is any such actual relationship or order between these entities or operations. Moreover, the term "include", "comprise" or any other variant is intended to cover non-exclusive inclusion, so that the article or device including a series of elements includes not only those elements, but also other elements that are not explicitly listed. In the absence of more restrictions, the elements defined by the sentence "including one..." do not exclude the existence of other identical elements in the article or device including the elements. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. The orientation or position relationship indicated by "up", "down", "left", "right", etc. is based on the orientation or position relationship shown in the accompanying drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation of the present invention.

[0037] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means 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 representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification.

[0038] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

Claims

1. A radar target multi-mode data fusion recognition method based on knowledge embedding model, characterized in that: include: Obtain radar echo data to be identified and corresponding domain knowledge data; Inputting the radar echo data to be identified and the corresponding domain knowledge data into a trained radar target recognition network for processing, extracting the range image feature tokens in the radar echo data to be identified, extracting the domain knowledge text feature tokens of the domain knowledge data, and fusing the range image feature tokens and the domain knowledge text feature tokens to obtain a fusion recognition result of the target to be identified; The trained radar target recognition network is obtained by training an initial radar target recognition network using preset category data as a training data set.

2. The radar target multi-mode data fusion recognition method based on the knowledge embedding model according to claim 1 is characterized in that: The trained radar target recognition network includes a trained first feature extraction module, a trained second feature extraction module and a trained mutual attention feature fusion recognition module; the radar echo data to be recognized and the corresponding domain knowledge data are input into the trained radar target recognition network for processing, the range image feature tokens in the radar echo data to be recognized are extracted, the domain knowledge text feature tokens of the domain knowledge data are extracted, and the range image feature tokens and the domain knowledge text feature tokens are fused and recognized to obtain the fusion recognition result of the target to be recognized, including: Using the trained first feature extraction module to extract features from the radar echo data to be identified, to obtain a range image feature token; Using the trained second feature extraction module to extract features from the domain knowledge data to obtain domain knowledge text feature tokens; The trained mutual attention feature fusion recognition module is used to perform feature fusion and recognition on the distance image feature token and the domain knowledge text feature token to obtain a fusion recognition result of the target to be recognized.

3. The radar target multi-mode data fusion recognition method based on knowledge embedding model according to claim 2 is characterized in that: The trained first feature extraction module includes a one-dimensional high-resolution range image preprocessing module and a range image feature extraction module; The step of using the trained first feature extraction module to extract features from the radar echo data to be identified to obtain a range profile feature token includes: The one-dimensional high-resolution range image preprocessing module is used to preprocess the radar echo data to be identified to obtain a processed one-dimensional range image; The range image feature extraction module is used to extract features from the processed one-dimensional range image to obtain a range image feature token.

4. The radar target multi-mode data fusion recognition method based on knowledge embedding model according to claim 2 is characterized in that: The trained second feature extraction module includes a domain knowledge text preprocessing module and a domain knowledge text feature extraction module; The step of using the trained second feature extraction module to extract features from the domain knowledge data to obtain domain knowledge text feature tokens includes: Using the domain knowledge text preprocessing module to preprocess the domain knowledge data to obtain processed domain knowledge text tokens; The domain knowledge text feature extraction module is used to extract features from the processed domain knowledge text tokens to obtain domain knowledge text feature tokens.

5. The radar target multi-mode data fusion recognition method based on knowledge embedding model according to claim 2 is characterized in that: The trained mutual attention feature fusion recognition module includes multiple cascaded mutual attention fusion modules and fusion classification layers, wherein each of the mutual attention fusion modules includes a first normalization layer, a second normalization layer, a mutual attention fusion layer, a third normalization layer, a fourth normalization layer, a first feedforward layer, and a second feedforward layer; the trained mutual attention feature fusion recognition module is used to perform feature fusion and recognition on the distance image feature token and the domain knowledge text feature token to obtain a fusion recognition result of the target to be recognized, including: Using the first normalization layer to normalize the range image feature token to obtain a first feature; Processing the first feature using the mutual attention fusion layer to obtain a second feature; Adding the second feature to the range image feature token to obtain a third feature; Normalizing the third feature using the third normalization layer to obtain a fourth feature; Processing the fourth feature using the first feed-forward layer to obtain a fifth feature; Adding the fifth feature to the third feature to obtain a one-dimensional high-resolution range image classification feature vector; The second normalization layer is used to normalize the domain knowledge text feature token to obtain a sixth feature; Processing the sixth feature using the mutual attention fusion layer to obtain a seventh feature; Adding the seventh feature to the domain knowledge text feature token to obtain an eighth feature; Normalizing the eighth feature using the fourth normalization layer to obtain a ninth feature; Using the second feed-forward layer to process the ninth feature to obtain a tenth feature; Adding the tenth feature to the eighth feature to obtain a domain knowledge text classification feature vector; Concatenate the one-dimensional high-resolution range image classification feature vector with the domain knowledge text classification feature vector to obtain a fused feature vector; The fusion classification layer is used to classify the fusion feature vector to obtain a fusion recognition result of the target to be recognized.

6. The radar target multi-mode data fusion recognition method based on knowledge embedding model according to claim 1 is characterized in that: The training process of the trained radar target recognition network includes: Acquire data of a plurality of preset categories to construct a training data set; wherein the training data set includes a plurality of samples, each sample including radar echo data and corresponding domain knowledge data; Input some samples in the training data set into the The radar target recognition network to be trained is trained to obtain the The prediction results of the classification output during the training process; According to The prediction results of the classification output during the training process are consistent with the The real label of the sample of the radar target recognition network to be trained is used to calculate the classification loss and serve as the The classification loss of the training process; at the same time, according to the The one-dimensional high-resolution range image classification feature vector output during the first training process is matched with the domain knowledge text classification feature vector to calculate the matching loss and use it as the first Matching loss of the training process; According to The classification loss and matching loss of the first training process are back-propagated to update the The network parameters of the radar target recognition network to be trained are obtained The radar target recognition network to be trained is iterated until the number of training times or the degree of convergence meets the preset conditions, thereby obtaining the trained radar target recognition network.

7. The radar target multi-mode data fusion identification method based on knowledge embedding model according to claim 6 is characterized in that: The mutual attention feature fusion recognition module in the initial radar target recognition network includes an initial mutual attention fusion module, an initial one-dimensional high-resolution range image classification layer, an initial domain knowledge text classification layer and an initial fusion classification layer; The process of obtaining the classification loss of the training process includes: The initial mutual attention fusion module is used to generate the One-dimensional high-resolution range image classification feature vector and domain knowledge text classification feature vector in the training process; The initial one-dimensional high-resolution range image classification layer is used to classify the first The one-dimensional high-resolution range image classification feature vector generated in the first training process is classified to obtain a first prediction result, and the first prediction result is compared with the first training process. The real labels of the samples of the radar target recognition network to be trained are compared to obtain the first classification loss; The initial fusion classification layer is used to The fused feature vector generated in the first training process is classified to obtain the second prediction result, and the second prediction result is compared with the first training process. The real labels of the samples of the radar target recognition network to be trained are compared to obtain the second classification loss; The initial domain knowledge text classification layer is used to The domain knowledge text classification feature vector generated in the first training process is classified to obtain a third prediction result, and the third prediction result is compared with the first training process. The real labels of the samples of the radar target recognition network to be trained are compared to obtain the third classification loss; The first classification loss, the second classification loss and the third classification loss are added together to obtain the The classification loss of the training process.

8. The radar target multi-mode data fusion identification method based on knowledge embedding model according to claim 7 is characterized in that: The expressions of the first classification loss, the second classification loss or the third classification loss are all: ； in, means a first category loss, a second category loss or a third category loss, represents the true label, Represents the prediction result, Indicates different modes, They represent the one-dimensional high-resolution range image modality, the domain knowledge text classification modality and the fusion modality respectively. represents the total number of categories, Represents the category index.

9. The radar target multi-mode data fusion identification method based on knowledge embedding model according to claim 6 is characterized in that: The expression of the matching loss is: ； in, represents the matching loss value, represents the one-dimensional high-resolution range image classification feature vector, Represents the domain knowledge text classification feature vector, represents transpose, Represents the L2 norm of a vector.

Images