Radar target multimode data robust identification method under resource limited condition

Abstract

The invention discloses a radar target multimode data robust identification method under a resource limited condition. The radar target multimode data robust identification method comprises the steps of inputting radar multimode data to be identified into a trained radar multimode data identification model; wherein the data modality comprises a high-resolution one-dimensional range profile modality, a narrowband modulation spectrum modality and a state information sequence modality; the model comprises a single-mode hybrid expert learning module and a multi-mode feature confidence fusion module. Evidence, uncertainty and category probability corresponding to each modal data of the radar are obtained according to features corresponding to each modal data of the radar obtained by the trained multi-modal feature confidence fusion module and the trained single-modal hybrid expert learning module, and confidence fusion is carried out; and obtaining an identification result according to the obtained target confidence fusion result. According to the method, the features of each mode of the radar data are fully mined, the accuracy of description of the radar multi-mode data is improved, and the reliability and accuracy of target identification are improved through confidence fusion.

Description

Technical Field

[0001] The present invention belongs to the field of radar technology, and in particular relates to a method for robustly identifying multi-mode data of radar targets under resource-constrained conditions. Background Art

[0002] The radar high-resolution one-dimensional range profile (HRRP) reflects the distribution of the main scattering centers of the target along the radar line of sight at a certain radar viewing angle, including the geometric characteristics and electromagnetic scattering characteristics of the main structure of the target, and is widely used in the field of radar target recognition. However, due to the limitation of system resources, HRRP has the problem of information loss due to insufficient bandwidth. In addition, due to the complexity of the actual application environment, HRRP may also face the problem of being affected by noise and interference, making it difficult to robustly identify targets. Some existing radar target recognition technologies integrate HRRP and Synthetic Aperture Radar (SAR) data, but do not take into account the description of the physical characteristics of the target by radar data of different modes, resulting in poor radar target recognition effect. Summary of the invention

[0003] In order to solve the above problems existing in the prior art, the present invention provides a method for robustly identifying multi-mode data of radar targets under resource-constrained conditions.

[0004] 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 method for robustly identifying radar target multi-mode data under resource-constrained conditions, the method comprising: Inputting the radar multimodal data to be identified into the trained radar multimodal data identification model; wherein the modes of the radar multimodal data to be identified include: high-resolution one-dimensional range image mode, narrow-band modulation spectrum mode and state information sequence mode; the radar multimodal data identification model includes a single-modal hybrid expert learning module and a multimodal feature confidence fusion module; Obtaining features corresponding to each modal data of the radar in the multimodal data of the radar to be identified according to the trained single-modal hybrid expert learning module; Obtaining evidence, uncertainty and category probability corresponding to each modal data of radar in the multimodal data of radar to be identified according to the trained multimodal feature confidence fusion module and the features corresponding to each modal data of radar in the multimodal data of radar to be identified; wherein the evidence corresponding to each modal data of radar represents the support degree for each category to which the multimodal data of radar to be identified belongs; According to the evidence, uncertainty and category probability corresponding to each modal data of the radar in the multimodal data of the radar to be identified, confidence fusion is performed to obtain a target confidence fusion result; The recognition result of the radar multimodal data to be recognized is obtained according to the target confidence fusion result.

[0005] Optionally, the training process of the radar multimodal data recognition model includes: The training process of the radar multimodal data recognition model includes: Constructing a training data set; wherein each sample in the training data set is labeled with a corresponding real category, and the modality of each sample includes: a high-resolution one-dimensional range image modality, a narrow-band modulation spectrum modality, and a state information sequence modality; The radar multimodal data recognition model is trained according to the training data set and a preset loss function until a preset condition is met, thereby obtaining the trained radar multimodal data recognition model.

[0006] Optionally, the preset loss function includes: load balancing loss; the load balancing loss is expressed as follows: ; in, Representing modality The corresponding load balancing loss is Indicates the mode The corresponding number of experts, Indicates that in the modal Assigned to the experts in the unimodal hybrid expert learning module The token ratio, Indicates that in the modal Assigned to the expert The average probability score.

[0007] Optionally, the preset loss function further includes: confidence fusion loss; the confidence fusion loss is expressed as follows: ; in, represents the confidence fusion loss, represents the cross entropy loss, represents the regularization loss, Representation sample Parameter vector of the Dirichlet distribution after confidence fusion.

[0008] Optionally, the cross entropy loss is expressed as follows: ; in, For the sample No. True labels in one-hot format, Representation sample Predicted to be The probability of the class, for dimensional polynomial function, represents the number of categories, express The sample described in No. The value of the class, Indicates is the class probability vector obtained by Dirichlet parameters, represents the digamma function, Indicates the sample The Dirichlet intensity of .

[0009] Optionally, the regularization loss is expressed as follows: ; in, represents a vector of all 1s, represents the Kullback-Leibler divergence, For The Dirichlet parameter vector after removing non-misleading evidence, For sample The true label in one-hot format, is the probability density function of the Dirichlet distribution, is the gamma function, express Middle Dirichlet parameter after removing non-misleading evidence.

[0010] Optionally, the target confidence fusion result is expressed as follows: ; in, represents the target confidence fusion result, represents the uncertainty of the high-resolution one-dimensional range image mode, represents the uncertainty of the narrowband modulation spectrum mode, represents the uncertainty of the state information sequence mode, represents the evidence of the high-resolution one-dimensional range image modality, represents evidence of the narrowband modulation spectrum modality, Evidence representing the modality of the state information sequence.

[0011] Optionally, obtaining the evidence, uncertainty and category probability corresponding to each modal data of the radar in the multimodal data of the radar to be identified according to the trained multimodal feature confidence fusion module and the features corresponding to each modal data of the radar in the multimodal data of the radar to be identified includes: The features corresponding to each modal data of the radar in the radar multimodal data to be identified are subjected to the linear layer and Softplus in the multimodal feature confidence fusion module to obtain the evidence corresponding to each modal data of the radar; The uncertainty and category probability corresponding to each radar modal data in the radar multimodal data to be identified are calculated according to the evidence corresponding to each radar modal data.

[0012] The technical solution provided by the embodiments of the present invention may have the following beneficial effects: The present invention comprehensively utilizes the radar high-resolution one-dimensional range image mode, narrow-band Doppler modulation spectrum mode, and state information sequence mode, fully mines the characteristics of radar data from multi-modal and multi-dimensional perspectives, and improves the accuracy of the attribute description of radar multi-modal data to be identified; the present invention also takes into account the modal quality uncertainty that may exist in radar multi-modal data, and uses a single-modal hybrid expert learning module in model structure design to enhance the processing capability of the model, so that different expert networks are selected for processing data of different modes, and a multi-modal feature confidence fusion module based on modal uncertainty is designed in the fusion strategy to perform confidence fusion on radar modal data, effectively improving the reliability and accuracy of radar multi-modal data fusion target recognition.

[0013] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a flow chart of a method for robustly identifying multi-mode data of radar targets under resource-constrained conditions provided by an embodiment of the present invention; Figure 2 is a schematic diagram of a single-modal hybrid expert learning module provided by an embodiment of the present invention; Figure 3 It is a schematic diagram of a multimodal feature confidence fusion module provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0015] 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.

[0016] Figure 1 is a flowchart of a method for robustly identifying multi-mode data of radar targets under resource-constrained conditions provided by an embodiment of the present invention. Figure 1 As shown, the method may include the following steps: S101. Inputting radar multimodal data to be identified into a trained radar multimodal data identification model; wherein the modalities of the radar multimodal data to be identified include: high-resolution one-dimensional range image modality, narrow-band modulation spectrum modality and state information sequence modality; the radar multimodal data identification model includes a single-modal hybrid expert learning module and a multimodal feature confidence fusion module.

[0017] For example, the radar multimodal data to be identified can be referred to as follows: the radar multimodal data to be identified can be three types of aircraft target data, consisting of data of aircraft A, aircraft B, and aircraft C. Each data includes a radar one-dimensional high-resolution range image, whose dimension is 1×256; a narrowband modulation spectrum, whose dimension is 1×1024; a state information sequence, which includes 10 moments, each moment includes target distance, azimuth, pitch, attitude, altitude, speed, radar cross-sectional area, and the dimension is 10×7.

[0018] S102, obtaining features corresponding to each modal data of radar in the radar multimodal data to be identified according to the trained single-modal hybrid expert learning module.

[0019] It is understandable that each modality has a corresponding single-modal hybrid expert learning module, so the single-modal hybrid expert learning module can include: a distance image single-modal hybrid expert learning module, a modulation spectrum single-modal hybrid expert learning module and a state information sequence single-modal hybrid expert learning module. The range image single-modal hybrid expert learning module includes: a range image preprocessing module, a range image feature extraction module and a range image feature hybrid expert learning module. The input of the range image single-modal hybrid expert learning module is a high-resolution one-dimensional range image, and the output is the extracted range image features and the range image hybrid expert load balancing loss; the modulation spectrum single-modal hybrid expert learning module includes: a modulation spectrum preprocessing module, a modulation spectrum feature extraction module and a modulation spectrum feature hybrid expert learning module. The input of the modulation spectrum single-modal hybrid expert learning module is a narrowband modulation spectrum, and the output is the extracted modulation spectrum features and the modulation spectrum hybrid expert load balancing loss; the state information sequence single-modal hybrid expert learning module includes: a state information sequence preprocessing module, a state information sequence feature extraction module and a state information sequence feature hybrid expert learning module. The input of the state information sequence single-modal hybrid expert learning module is a state information sequence, and the output is the extracted state information sequence features and the state information sequence hybrid expert load balancing loss.

[0020] Specifically, Figure 2 is a schematic diagram of a single-mode hybrid expert learning module provided by an embodiment of the present invention, such as Figure 2As shown in the figure, taking the radar multimodal data to be identified as an example, the high-resolution one-dimensional range image in the radar multimodal data to be identified is input into the range image preprocessing module. The preprocessing can use the L2 norm for normalization and centroid alignment. Then it is input into the range image feature extraction module. The feature extraction can use a three-layer one-dimensional convolutional network, each layer contains 1 convolution layer, 1 batch normalization layer and 1 ReLU activation layer, the number of convolution channels is (8, 16, 32), the convolution kernel size is 3, the step size is 2, and the padding is 1; then the features extracted by the range image feature extraction module are input into the range image feature hybrid expert learning module. The range image feature hybrid expert learning module first adds a learnable parameter of the same length as the range image feature token sequence and a dimension of 128 as the position code. After the range image feature token sequence is added to the position code, it is passed through 1 Self-attention layers (including 3 linear layers with a node number of (256, 128)), 1 layer normalization layer, 1 expert probability mapping layer (including 1 linear layer with a node number of (128, 4)), 1 Softmax layer, 1 mixed expert layer (including 4 feedforward experts, each expert including 1 linear layer with a node number of (128, 512)), 1 GeLU activation layer, 1 linear layer (with a node number of (512, 128)), 1 layer normalization layer and an averaging layer are used to obtain the distance image feature output with a node number of (1, 128). The number of experts selected in this embodiment is 1.

[0021] The narrowband modulation spectrum in the radar multimodal data to be identified is input into the modulation spectrum preprocessing module. The preprocessing adopts denoising, translation of the main component and normalization, and then input into the modulation spectrum feature extraction module. The feature extraction adopts a three-layer one-dimensional convolutional network, each layer contains one convolution layer, one batch normalization layer, one ReLU activation layer and one maximum pooling layer. The number of convolution channels is (8, 16, 32), the convolution kernel size is 3, the step size is 2, the padding is 1, and the pooling step size is 2. Then the features extracted by the modulation spectrum feature extraction module are input into the modulation spectrum feature hybrid expert learning module. The modulation spectrum feature hybrid expert learning module first adds a learnable parameter with the same length as the modulation spectrum feature token sequence and a dimension of 128 as the bit After the position coding, the modulation spectrum feature token sequence is added to the position coding, and then passes through 1 self-attention layer (including 3 linear layers, the number of nodes is (128, 128)), 1 layer normalization layer, 1 expert probability mapping layer (including 1 linear layer, the number of nodes is (128, 4)), 1 Softmax layer, 1 mixed expert layer (including 4 feedforward experts, each expert contains 1 linear layer, the number of nodes is (128, 512)), 1 GeLU activation layer, 1 linear layer (the number of nodes is (512, 128)), 1 layer normalization layer, and an averaging layer to obtain the modulation spectrum feature output (the number of nodes is (1, 128)). The number of experts selected in this embodiment is 1.

[0022] The state information sequence in the radar multimodal data to be identified is input into the state information preprocessing module, the state information sequence is rearranged and normalized by data format, and then input into the state information feature extraction module. The feature extraction adopts a three-layer bidirectional long short-term memory network, each layer contains a forward long short-term memory layer, a reverse long short-term memory layer, and the hidden layer dimension is 64. Then it is input into the state information feature hybrid expert learning module. The state information feature hybrid expert learning module first adds a learnable parameter with a dimension of 128 and the same length as the token of the state information feature sequence as the position code. After the token sequence of the state information feature sequence is added to the position code, , after 1 self-attention layer (including 3 linear layers, the number of nodes is (128, 128)), after 1 layer normalization layer, after 1 expert probability mapping layer (including 1 linear layer, the number of nodes is (128, 4)), 1 Softmax layer, 1 mixed expert layer (including 4 feedforward experts, each expert includes 1 linear layer, the number of nodes is (128, 512)), 1 GeLU activation layer, 1 linear layer (the number of nodes is (512, 128)), after 1 layer normalization layer, after an averaging layer, the state information feature output (the number of nodes is (1, 128)) is obtained. The number of experts selected in this embodiment is 1.

[0023] S103, obtaining evidence, uncertainty and category probability corresponding to each radar modal data in the radar multimodal data to be identified according to the trained multimodal feature confidence fusion module and the features corresponding to each radar modal data in the radar multimodal data to be identified; wherein the evidence corresponding to each radar modal data represents the support level for each category to which the radar multimodal data to be identified belongs.

[0024] It can be understood that the uncertainty corresponding to each radar modal data in the radar multimodal data to be identified represents the reliability of the result; the category probability corresponding to each radar modal data in the radar multimodal data to be identified represents the confidence of the identification result of the category to which each radar modal data in the radar multimodal data to be identified belongs.

[0025] Optionally, Figure 3 is a schematic diagram of a multimodal feature confidence fusion module provided by an embodiment of the present invention, and S103 may include: The features corresponding to each modal data of the radar in the radar multimodal data to be identified are processed through the linear layer and Softplus in the multimodal feature confidence fusion module to obtain the evidence corresponding to each modal data of the radar; The uncertainty and category probability corresponding to each radar modal data in the radar multimodal data to be identified are calculated based on the evidence corresponding to each radar modal data.

[0026] It can be understood that the features corresponding to each modal data of the radar in the radar multimodal data to be identified are respectively subjected to the linear layer and Softplus to obtain the evidence corresponding to each modal data of the radar. The number of nodes in the three linear layers is (128, 3). The Dirichlet distribution parameters, category beliefs, uncertainty, and category probabilities are calculated. In this embodiment, the number of categories is ,For example It can be 3, the category belief corresponding to each mode , uncertainty and class probabilities , calculated as follows: ; ; ; ; ; Among them, the mode , represents the high-resolution one-dimensional range image mode, represents the narrowband modulation spectrum mode, Represents the state information sequence mode, For modal The Dirichlet distribution parameter is For modal The Dirichlet intensity of For modal evidence.

[0027] S104, performing confidence fusion according to the evidence, uncertainty and category probability corresponding to each modal data of the radar in the radar multimodal data to be identified, and obtaining a target confidence fusion result.

[0028] Optionally, the target confidence fusion result is expressed as follows: ; in, represents the target confidence fusion result, represents the uncertainty of the high-resolution one-dimensional range image mode, represents the uncertainty of the narrowband modulation spectrum mode, represents the uncertainty of the state information sequence mode, This indicates evidence for the high-resolution one-dimensional range profile mode, shows evidence of a narrowband modulation spectrum mode, Evidence for representing sequential modality of state information.

[0029] S105. Obtaining recognition results of the radar multimodal data to be recognized according to the target confidence fusion results.

[0030] It can be understood that the confidence that the radar multimodal data to be identified belongs to each category is obtained according to the target confidence fusion result, and the category with the highest confidence is used as the category of the radar multimodal data to be identified, that is, the identification result of the radar multimodal data to be identified.

[0031] Optionally, the training process of the radar multimodal data recognition model includes: Constructing a training data set; wherein each sample in the training data set is labeled with a corresponding real category, and the modalities of each sample include: high-resolution one-dimensional range image modality, narrow-band modulation spectrum modality, and state information sequence modality; The radar multimodal data recognition model is trained according to the training data set and the preset loss function until the preset conditions are met, thereby obtaining a trained radar multimodal data recognition model.

[0032] For example, the total number of samples in the training data set can be 3000, each sample contains 1 radar high-resolution one-dimensional range image, 1 narrowband modulation spectrum and 1 target state information sequence, each target state information sequence contains data at 10 moments, and each moment contains target distance, azimuth, pitch, attitude, altitude, speed, and radar cross-sectional area. The training data set contains 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 samples in the test data set is 1500, and the composition of each sample is the same as that of the training data set, including 500 samples of aircraft A, 500 samples of aircraft B, and 500 samples of aircraft C. The test data set is used to test the performance of the radar multimodal data recognition model. The preset condition is that the sum of each loss is minimized during a finite number of iterative training processes. For example, when training the radar multimodal data recognition model, the batch size in each round can be 64, the initial learning rate is 1e-4, the optimizer is AdamW, and the total number of iterations is set to 100.

[0033] Optionally, the preset loss function includes: load balancing loss; the load balancing loss is expressed as follows: ; in, Representing modality The corresponding load balancing loss is Representing modality The corresponding number of experts, Indicates that in modal Assigned to the experts in the unimodal hybrid expert learning module The token ratio, Indicates that in modal Assigned to experts The average probability score.

[0034] ; in, Representing modality The corresponding number of feature tokens, is the indicator function, represents any batch of training data in the training dataset, Representing modality training data, According to The obtained probability of selecting each expert; ; in, According to Get the selection expert probability.

[0035] Optionally, the preset loss function further includes: confidence fusion loss; the confidence fusion loss is expressed as follows: ; in, represents the confidence fusion loss, represents the cross entropy loss, represents the regularization loss, Representation sample Parameter vector of the Dirichlet distribution after confidence fusion.

[0036] Optionally, the cross entropy loss is expressed as follows: ; in, For sample No. True labels in one-hot format, Representation sample Predicted to be The probability of the class, for dimensional polynomial function, represents the number of categories, express Medium Sample No. The value of the class, Indicates is the class probability vector obtained by Dirichlet parameters, represents the digamma function, Representation sample The Dirichlet intensity of .

[0037] The regularization loss is expressed as follows: ; in, represents a vector of all 1s, represents the Kullback-Leibler divergence, For The Dirichlet parameter vector after removing non-misleading evidence, For sample The true label in one-hot format, is the probability density function of the Dirichlet distribution, is the gamma function, express Middle Dirichlet parameter after removing non-misleading evidence.

[0038] The effectiveness of the present invention is illustrated by simulation experiments below.

[0039] Specifically, the hardware platform of the simulation experiment 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. The software platform of the simulation experiment is: Python3.8, Pytorch 2.0.

[0040] In the simulation experiment of the present invention, the present invention and a prior art are used to conduct target recognition experiments on three types of aircraft, and the recognition rates are compared. In order to simulate the disturbance in the actual scene, a Gaussian perturbation with a standard deviation of 1 is added to any mode of 50% of the test data sets in the experiment. The prior art refers to the paper "Deep Residual Learning for Image Recognition", which applies the ResNet network to the recognition model. The original network is designed for two-dimensional image data. In this embodiment, its main structure is retained and the two-dimensional operation is modified into a one-dimensional operation to adapt to HRRP 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 identified samples to the total number of samples in the test data set. The higher the recognition rate, the more correctly identified samples there are, and the better the recognition performance of the model.

[0041] Table 1

[0042] 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 in disturbance scenarios.

[0043] The present invention comprehensively utilizes the radar high-resolution one-dimensional range image mode, narrow-band Doppler modulation spectrum mode, and state information sequence mode, fully mines the characteristics of radar data from multi-modal and multi-dimensional perspectives, and improves the accuracy of the attribute description of radar multi-modal data to be identified; the present invention also takes into account the modal quality uncertainty that may exist in radar multi-modal data, and uses a single-modal hybrid expert learning module in model structure design to enhance the processing capability of the model, so that different expert networks are selected for processing data of different modes, and a multi-modal feature confidence fusion module based on modal uncertainty is designed in the fusion strategy to perform confidence fusion on radar modal data, effectively improving the reliability and accuracy of radar multi-modal data fusion target recognition.

[0044] 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.

[0045] Although the present invention is described herein in conjunction with various embodiments, in the process of implementing the claimed invention, those skilled in the art can understand and implement other changes to the disclosed embodiments by viewing the drawings and the disclosed content. In the description of the present invention, the term "comprising" does not exclude other components or steps, "one" or "an" does not exclude multiple situations, and "multiple" means two or more, unless otherwise clearly and specifically limited. In addition, certain measures are recorded in different embodiments, but this does not mean that these measures cannot be combined to produce good results.

[0046] 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 robust identification method for radar target multi-mode data under resource-constrained conditions, characterized in that: The method comprises: Inputting the radar multimodal data to be identified into the trained radar multimodal data identification model; wherein the modes of the radar multimodal data to be identified include: high-resolution one-dimensional range image mode, narrow-band modulation spectrum mode and state information sequence mode; the radar multimodal data identification model includes a single-modal hybrid expert learning module and a multimodal feature confidence fusion module; Obtaining features corresponding to each modal data of the radar in the multimodal data of the radar to be identified according to the trained single-modal hybrid expert learning module; Obtaining evidence, uncertainty and category probability corresponding to each modal data of radar in the multimodal data of radar to be identified according to the trained multimodal feature confidence fusion module and the features corresponding to each modal data of radar in the multimodal data of radar to be identified; wherein the evidence corresponding to each modal data of radar represents the support degree for each category to which the multimodal data of radar to be identified belongs; According to the evidence, uncertainty and category probability corresponding to each modal data of the radar in the multimodal data of the radar to be identified, confidence fusion is performed to obtain a target confidence fusion result; The recognition result of the radar multimodal data to be recognized is obtained according to the target confidence fusion result.

2. The method for robust identification of radar target multi-mode data under resource-constrained conditions according to claim 1 is characterized in that: The training process of the radar multimodal data recognition model includes: Constructing a training data set; wherein each sample in the training data set is labeled with a corresponding real category, and the modality of each sample includes: a high-resolution one-dimensional range image modality, a narrow-band modulation spectrum modality, and a state information sequence modality; The radar multimodal data recognition model is trained according to the training data set and a preset loss function until a preset condition is met, thereby obtaining the trained radar multimodal data recognition model.

3. The method for robust identification of radar target multi-mode data under resource-constrained conditions according to claim 2 is characterized in that: The preset loss function includes: load balancing loss; the load balancing loss is expressed as follows: ； in, Representing modality The corresponding load balancing loss is Indicates the mode The corresponding number of experts, Indicates that in the modal Assigned to the experts in the unimodal hybrid expert learning module The token ratio, Indicates that in the modal Assigned to the expert The average probability score.

4. The method for robust identification of radar target multi-mode data under resource-constrained conditions according to claim 2, characterized in that: The preset loss function also includes: confidence fusion loss; the confidence fusion loss is expressed as follows: ； in, represents the confidence fusion loss, represents the cross entropy loss, represents the regularization loss, Representation sample Parameter vector of the Dirichlet distribution after confidence fusion.

5. The method for robust identification of radar target multi-mode data under resource-constrained conditions according to claim 4 is characterized in that: The cross entropy loss is expressed as follows: ； in, For the sample No. True labels in one-hot format, Representation sample Predicted to be The probability of the class, for dimensional polynomial function, represents the number of categories, express The sample described in No. The value of the class, Indicates is the class probability vector obtained by Dirichlet parameters, represents the digamma function, Indicates the sample The Dirichlet intensity of .

6. The method for robust identification of radar target multi-mode data under resource-constrained conditions according to claim 5, characterized in that: The regularization loss is expressed as follows: ； in, represents a vector of all 1s, represents the Kullback-Leibler divergence, For The Dirichlet parameter vector after removing non-misleading evidence, For sample The true label in one-hot format, is the probability density function of the Dirichlet distribution, is the gamma function, express Middle Dirichlet parameter after removing non-misleading evidence.

7. The method for robust identification of radar target multi-mode data under resource-constrained conditions according to claim 1, characterized in that: The target confidence fusion result is expressed as follows: ； in, represents the target confidence fusion result, represents the uncertainty of the high-resolution one-dimensional range image mode, represents the uncertainty of the narrowband modulation spectrum mode, represents the uncertainty of the state information sequence mode, represents the evidence of the high-resolution one-dimensional range image modality, represents evidence of the narrowband modulation spectrum modality, Evidence representing the modality of the state information sequence.

8. The method for robust identification of radar target multi-mode data under resource-constrained conditions according to claim 1, characterized in that: The obtaining of evidence, uncertainty and category probability corresponding to each modal data of the radar in the multimodal data of the radar to be identified based on the trained multimodal feature confidence fusion module and the features corresponding to each modal data of the radar in the multimodal data of the radar to be identified includes: The features corresponding to each modal data of the radar in the radar multimodal data to be identified are subjected to the linear layer and Softplus in the multimodal feature confidence fusion module to obtain the evidence corresponding to each modal data of the radar; The uncertainty and category probability corresponding to each radar modal data in the radar multimodal data to be identified are calculated according to the evidence corresponding to each radar modal data.

Images