Space micro-motion target recognition method based on multi-band data feature and text information fusion

Abstract

The application discloses a space micro-motion target recognition method based on multi-band data features and text information fusion, comprising: dividing narrowband radar scattering cross-section area data, wideband high-resolution range image data and text data of space micro-motion targets under multiple frequency bands to generate a training set and a test set; constructing a target recognition network comprising a feature extraction module, a multi-band graph attention module and a classifier; inputting the training set into the target recognition network, training through a back propagation algorithm, testing by using the test set, and obtaining a target recognition network meeting the conditions; and deploying the network on a device to recognize space micro-motion targets in real time. The application fully excavates complementary information of different frequency bands and different modal data by extracting multi-band multi-modal features and dynamically fusing by using a graph attention mechanism, significantly improves the recognition accuracy of space micro-motion targets, and can be widely applied to ground-based wideband and narrowband radar systems.

Description

Technical Field

[0001] This application belongs to the field of spatial micro-movement target recognition technology, specifically involving a spatial micro-movement target recognition method based on the fusion of multi-band data features and text information. Background Technology

[0002] In the field of radar target identification, the accurate identification of space targets with minute movements (such as targets undergoing spin, precession, or tumbling) is one of the technical challenges. Minor movements refer to the tiny vibrations or rotational motions of a target or its components relative to the radar line of sight, excluding the main body motion. When radar observes space targets with minute movements using alternating wide and narrow band modes, it can simultaneously obtain both narrow-band and wide-band echoes of the target. The target's narrow-band radar cross section (RCS) sequence reflects the overall scattering intensity of the target over time, while the wide-band high-resolution range profile (HRRP) effectively characterizes the target's structural characteristics and size information. Therefore, fusing multi-mode radar data from different frequency bands can theoretically significantly improve the accuracy of identifying space targets with minute movements.

[0003] To address the aforementioned issues, researchers have proposed several technical solutions. For example, Penghui Wang et al., in their paper "Intelligent radar HRRP target recognition based on CNN-BERT model," constructed a Convolutional Neural Network-Bidirectional Encoder Representations from Transformers (CNN-BERT) model for feature extraction and target recognition of HRRP data. This method first performs normalization and translation alignment preprocessing on the HRRP data, then uses a convolutional embedding module to extract the local spatial structure features of the HRRP and generate sequence representations. Next, it models the long-range temporal dependencies between range units using a multi-head self-attention mechanism in the BERT module, and finally outputs the recognition results through attention weighting and a Softmax classifier. As another example, Haozhe Qiu et al., in their paper "Feature-Level Fusion Recognition of Radar Targets Based on Regularized Sequential Representation Network," proposed a radar target feature-level fusion recognition method based on a Regularized Sequential Representation Network (RSRNet). This method uses RCS sequence, HRRP sequence, and time spectrum as three different modal inputs, and performs initial feature extraction through convolutional neural networks. Then, a sequence fusion mechanism is constructed to perform cross-modal feature fusion step by step in the order of "RCS→time spectrum→HRRP". A dual-branch pixel attention fusion unit is designed to realize adaptive weighting of different modal features. Finally, the fused features are input into a classifier to complete target recognition.

[0004] However, the aforementioned existing technologies still have the following shortcomings. First, most existing methods only utilize single-band (such as narrowband or broadband) echoes for single-mode feature extraction, losing the separable features contained in wideband and narrowband multimodal data, and ignoring important discriminative information contained in the RCS spectrum and semantic information such as target physical characteristics and structural attributes contained in text descriptions, resulting in limited recognition performance. Second, most existing technologies focus on single-station radar observation architectures, which are limited by the inherent azimuth constraints and environmental noise of a single station, resulting in viewpoint sensitivity and information incompleteness in the acquired echo data. More importantly, existing methods are all based on feature extraction and fusion only on data from a single frequency band. Especially when the separability difference between different modal data is large, existing methods lack effective mechanisms for mining and fusing separable features, thus limiting further improvement in target recognition performance. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this application provides a spatial micro-motion target recognition method based on the fusion of multi-band data features and text information. The technical problem to be solved by this application is achieved through the following technical solution: A spatial micro-movement target recognition method based on the fusion of multi-band data features and text information includes: S1: Divide historical narrowband RCS data, broadband HRRP data and text data of spatial micro-movement targets in multiple frequency bands to generate training set and test set; S2: Construct a target recognition network that fuses multi-band data features and text information. The target recognition network includes a feature extraction module, a multi-band graph attention module, and a classifier. S3: Input the training set into the multi-band data feature and text information fusion recognition network, train it through the backpropagation algorithm, and then use the test set to test the trained target recognition network to obtain a target recognition network that meets the conditions. S4: Deploy the qualified target recognition network on the device to perform real-time recognition of spatial micro-moving targets.

[0006] Beneficial effects: 1. This application extracts multimodal features from narrowband RCS data, broadband HRRP data and text information of spatial micro-moving targets in multiple frequency bands, and uses a time-domain and frequency-domain dynamic fusion module and a broadband HRRP feature extractor to perform deep feature mining on each modality of data. This overcomes the problem of information loss caused by the use of only single-modal features in the prior art and significantly improves the recognition accuracy of spatial micro-moving targets.

[0007] 2. This application designs a multi-band graph attention module, which uses narrowband RCS fusion features, broadband HRRP features, and text features from multiple frequency bands as nodes in a graph structure. A graph attention network is used to achieve dynamic weighted fusion of cross-band and cross-modal features. Compared to existing technologies that only utilize single-band data for identification, this application fully leverages the complementary information of multi-band echo data, effectively overcoming the inherent azimuth constraints of single-station radar and the viewpoint sensitivity and information incompleteness caused by environmental noise, significantly improving the robustness of identifying small-moving targets in space.

[0008] 3. This application introduces textual information into the spatial micro-motion target recognition task. A pre-trained encoder semantically encodes the textual description containing the target's physical characteristics and structural attributes, and then fuses it with radar echo features in a multimodal manner. The prior semantic constraints provided by the textual information effectively compensate for the shortcomings of radar data in representing target attributes, further enhancing the discriminative power of the fused features.

[0009] 4. This application designs a time-domain and frequency-domain dynamic fusion module in the target narrowband RCS feature extraction. Through multiple time-domain and frequency-domain feature interaction layers and a dynamic weighting mechanism, it achieves deep coupling and adaptive fusion of time-domain and frequency-domain features. In the multi-band graph attention module, the graph attention network adaptively learns the attention weights between features of different modalities and frequency bands, achieving interpretable feature fusion. This dynamic fusion mechanism is particularly suitable for complex scenarios with large differences in the separability of data from different modalities, ensuring the stability of recognition performance.

[0010] 5. The target recognition network constructed in this application has a clear structure and modular design. After training, it can be easily deployed in radar signal processing equipment or edge computing equipment to quickly process and recognize multi-frequency and multi-modal data received in real time. It has good engineering application prospects and can be widely used in multiple ground-based wide and narrow band radar systems.

[0011] The present application will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0012] Figure 1 This is a flowchart illustrating a spatial micro-movement target recognition method based on the fusion of multi-band data features and text information provided in this application; Figure 2 This is a schematic diagram of the target recognition network provided in this application; Figure 3 This is a schematic diagram of the structure of the 1-D CNN provided in this application; Figure 4 This is a schematic diagram of the structure of the 1-D TCN provided in this application; Figure 5This is a schematic diagram of the TFDF provided in this application; Figure 6 This is a schematic diagram of the structure of the HFE provided in this application. Detailed Implementation

[0013] The present application will be described in further detail below with reference to specific embodiments, but the implementation of the present application is not limited thereto.

[0014] like Figure 1 As shown, this application provides a spatial micro-movement target recognition method based on the fusion of multi-band data features and text information, including: S1: Divide historical narrowband RCS data, broadband HRRP data and text data of spatial micro-movement targets in multiple frequency bands to generate training set and test set; This application acquires narrowband RCS data, broadband HRRP data, and text data of space micro-moving targets in the L-band, S-band, and X-band. Specifically, the initial elevation angle ranges from 31° to 43° in the L-band, 21° to 33° in the S-band, and 41° to 53° in the X-band. The initial elevation angle varies uniformly in 1° intervals, the precession frequency varies uniformly in 0.2Hz intervals, and the precession angle varies uniformly in 0.15° intervals. For each type of target in each band, the data is obtained... 1573 sets of narrowband echo data were obtained, and Gaussian noise was added to obtain data with signal-to-noise ratios of 5dB, 10dB, 15dB, and 20dB. Samples corresponding to initial elevation angles of 21° to 30°, 31° to 40°, and 41° to 50° were taken as training sets, resulting in 4840 samples for the four target classes in each frequency band. Samples corresponding to initial elevation angles of 31° to 33°, 41° to 43°, and 51° to 53° were taken as test sets, resulting in 1452 samples for the four target classes in each frequency band.

[0015] S2: Construct a target recognition network that fuses multi-band data features and text information. The target recognition network includes a feature extraction module, a multi-band graph attention module, and a classifier. refer to Figure 2The feature extraction module includes a target narrowband RCS feature extraction branch, a broadband HRRP feature extraction branch, and a text feature extraction branch. The target RCS feature extraction branch is constructed by sequentially connecting a 1-D Temporal Convolutional Network (1-D TCN), a 1-D Convolutional Neural Network (1-D CNN), and a Time and Frequency Feature Dynamic Fusion (TFDF) module. The 1-D TCN extracts the temporal features of the target narrowband RCS data, and the 1-D CNN extracts the frequency features. The input of the TFDF module is connected to the outputs of both the 1-D TCN and the 1-D CNN, fusing the temporal and frequency features to output fused narrowband RCS features. The wideband HRRP feature extraction branch is constructed by a wideband HRRP Feature Extractor (HFE) to extract wideband HRRP features and output wideband HRRP features. The text feature extraction branch is constructed by a pre-trained Tiny-Bert encoder to encode text information and output text features. The input of the multi-band graph attention module is connected to the target narrowband RCS fusion feature, broadband HRRP feature and text feature output by the feature extraction module in each frequency band, respectively, and is used to dynamically fuse features of different frequency bands and different modalities to output the final fused feature. The input of the classifier is connected to the output of the multi-band graph attention module, and is used to output a multimodal classification result based on the final fused features.

[0016] This application extracts RCS frequency domain features using a 1-D CNN. The 1-D CNN includes a one-dimensional convolutional layer, a batch normalization layer, a Mish activation function, and a one-dimensional max pooling layer connected in sequence. The one-dimensional convolutional layer has a kernel size of 3, a stride of 1, a padding size of 1, and 16 feature mapping channels. The one-dimensional max pooling layer has a kernel size of 2 and a stride of 2.

[0017] refer to Figure 3In the 1-D CNN, "Conv16k3 / s1 / p1" indicates a one-dimensional convolution operation with 16 feature mapping channels, a kernel size of 3, a stride of 1, and padding of 1. "BN / Mish" indicates the use of batch normalization and the Mish activation function. "MaxPoolk2 / s2" indicates a one-dimensional max pooling operation with a kernel size of 2 and a stride of 2. The number of nodes in the fully connected layer is also set to 128.

[0018] The 1-D TCN includes a one-dimensional temporal convolutional layer, a weight normalization layer, a ReLU activation function, and a residual connection operation connected in sequence; wherein, the kernel size of the one-dimensional temporal convolutional layer is 2, the stride is 1, the dilation factor is 1, and the number of feature mapping channels is 16. refer to Figure 4 In 1-D TCN, "Conv16k2 / s1 / d1" indicates a one-dimensional temporal convolution operation with 16 feature mapping channels, a kernel size of 2, a stride of 1, and an inflation factor of 1. "WN / ReLU" indicates the use of weight normalization and the ReLU activation function, with the number of nodes in the fully connected layer set to 128.

[0019] refer to Figure 5 The TFDF module contains multiple time-domain and frequency-domain feature interaction layers, each consisting of a time-domain and frequency-domain coupling block. The time-domain and frequency-domain coupling block processes time-domain and frequency-domain features in parallel. The specific processing procedure is as follows: The input time-domain and frequency-domain features are split, and then linear transformations are performed on the discretized state transition matrix and projection matrix of the corresponding modes to obtain intermediate time-domain and intermediate frequency-domain features. The intermediate time-domain features and intermediate frequency-domain features are concatenated, then mapped to a specified dimension through a 1-D CNN. The residual connections are then added to the input time-domain features and frequency-domain features respectively to obtain the output time-domain features and output frequency-domain features of this layer. The output time-domain features and output frequency-domain features are used as the input to the next layer's time-frequency coupling block.

[0020] For time domain features and frequency domain features First, it undergoes preprocessing; the calculation process is as follows:

[0021] in, This indicates a splitting operation. After being input into the TFDF module, the data fusion features are obtained, namely the time-domain and frequency-domain fusion features. , This indicates the dimension of data fusion features. The TFDF module contains... Each layer consists of a time-domain and frequency-domain feature interaction layer, with each layer composed of a time-and-frequency coupled (TFC) block. Within each TFC block, time-domain and frequency-domain features are processed in parallel. Simultaneously, the dynamically fused features are added to the processed time-domain and frequency-domain features and used as input for the next layer. The calculation process is as follows: ; Among them, through Obtain intermediate features and The calculation process can be expressed as: ; in, This represents the hidden state of the corresponding modality. This represents the discretized state transition matrix for the corresponding mode. and This represents the projection matrix of the corresponding mode. This represents the hidden state dimension. and After concatenation and mapping to an appropriate dimension using a 1-D CNN, the output features of each layer are obtained through residual connections. and At this point, they all carry features from the other modalities and are used as new input features to input the next layer of TFC blocks. The calculation process is as follows: ; Continue to refer to Figure 5 The TFDF module further includes a pooling layer, a 1-D CNN, and a weighted layer connected in sequence; The input of the pooling layer is connected to the output of the last time-domain and frequency-domain feature interaction layer, and is used to perform max pooling operations on the final output time-domain features and frequency-domain features respectively to obtain pooled time-domain features and pooled frequency-domain features; the input of the one-dimensional convolutional neural network is connected to the output of the pooling layer, and is used to concatenate the pooled time-domain features and pooled frequency-domain features and map them to obtain dynamic weights; the input of the weighting layer is connected to the output of the pooling layer and the output of the one-dimensional convolutional neural network respectively, and is used to perform weighted summation on the pooled time-domain features and pooled frequency-domain features using the dynamic weights to output the target narrowband RCS fusion feature.

[0022] After feature interaction, this application uses max pooling and dynamic weighted fusion to obtain the final fused features. The calculation process is as follows: ; in, and It is a dynamic weight obtained by concatenating time-domain and frequency-domain features and mapping them through a simple 1-D CNN network. The concatenation and mapping, rather than mapping them separately, is to allow the network to capture the interaction information between modalities.

[0023] refer to Figure 6 The HFE comprises a linear layer, a convolutional layer, an activation function, a state-space model layer, and a multi-layer perceptron (MLP) connected in sequence. The input high-resolution range image sequence first undergoes a linear transformation through the linear layer, and then is processed by the convolutional layer and activation function to obtain initial features. The state-space model layer is used to update the state and calculate the output of the initial features. Internally, it uses a discretized state-space parameter matrix to recursively calculate the hidden states to obtain the state-space output features. The MLP is used to perform nonlinear mapping on the state-space output features to output high-resolution range image features.

[0024] Broadband HRRP features were extracted using HFE, and the HRRP sequence was analyzed. The calculation process in the input resolution range image feature extractor is as follows: ; Features processed by linear layers, convolutions, and activation functions Intermediate features are obtained through a State-Space Model (SSM). The calculation process can be expressed as follows: ; Then, HRRP features were obtained through MLP. The calculation process can be expressed as follows: .

[0025] Text features are obtained by encoding text information using a pre-trained Tiny-Bert encoder. .

[0026] Continue to refer to Figure 2 The multi-band graph attention module is specifically used for: Target narrowband RCS fusion features, broadband HRRP features, and text features under multiple frequency bands are used as child nodes. The child nodes of different frequency bands are used to obtain multiple parent nodes through MLP, thereby constructing a graph structure. In this structure, there are no edges connecting child nodes, but there are edges connecting parent nodes. Child nodes are only connected to the parent nodes of the corresponding frequency band. The nodes in the graph structure are propagated and aggregated through a graph attention network to obtain multimodal fusion features under each frequency band. The multimodal fusion features under each frequency band are dynamically weighted to obtain the final fusion feature.

[0027] This application uses RCS features, HRRP features, and text features in the L, S, and X frequency bands as child nodes. ,in Indicates the node frequency band. To represent the node modes, child nodes of different frequency bands are used to obtain three parent nodes. ; in, This indicates a splicing operation. express The parent node of the frequency band. It's important to note that there are no edges connecting child nodes, but edges connect parent nodes. A child node is only connected to the parent node of its corresponding frequency band. This is included during propagation in MGAT. Layer, for any node in the constructed graph , No. Hidden layer representation The calculation formula is: ; in, , Indicates the first Layer nodes Relative to node Attention value, and These are the learnable weight vector and the weight matrix, respectively. Represents the nodes in the graph The set of adjacent nodes, This represents the activation function. After GAT, the aggregated features of the three parent nodes are obtained. , and That is, the multimodal fusion features under various frequency bands, among which, After obtaining the features of each frequency band, the final fused features are obtained through dynamic weighting. .

[0028] S3: Input the training set into the multi-band data feature and text information fusion recognition network, train it through the backpropagation algorithm, and then use the test set to test the trained target recognition network to obtain a target recognition network that meets the conditions. The classifier used in this application employs a Softmax classifier, with a cross-entropy loss function. This cross-entropy loss is calculated based on the one-hot encoding of the sample's true label and the probability output by the Softmax classification function. Backpropagation is used when training the target recognition network, and the network parameters are adjusted according to the direction of decrease in the overall loss function. The overall loss function is the product of the alignment loss and its coefficient, plus the product of the recognition loss and its coefficient. The recognition loss is the cross-entropy loss. The alignment loss is used to constrain the consistency of the distribution of different modal features in the embedding space.

[0029] In MDT-FR Net, a Softmax classifier is used to obtain the final multimodal classification result, where the loss function of the classifier part is: ; in, The output of the Softmax classification function. For the true labels of the samples, Let be the total number of target categories. Combining the loss functions of each part of the MDT-FR Net, the overall network loss function is: ; in, For alignment loss coefficients, To identify the loss coefficient.

[0030] This application inputs the training set into a target recognition network that fuses multi-band data features and text information, and trains it using the backpropagation algorithm; the test set is input into the trained multi-band data feature and text information fusion recognition network for testing, and the target recognition network that meets the conditions is selected.

[0031] S4: Deploy the qualified target recognition network on the device to perform real-time recognition of spatial micro-moving targets.

[0032] After training and testing the target recognition network that integrates multi-band data features and text information, and obtaining a target recognition network that meets the conditions (i.e., the recognition performance meets the preset requirements, such as achieving the best recognition accuracy on the test set), the network is deployed in radar signal processing equipment or edge computing equipment for real-time recognition of spatial micro-moving targets.

[0033] The deployment process specifically includes: solidifying and loading the network structure parameters and weight parameters of the trained target recognition network into the device's processor. In practical applications, the device receives narrowband RCS data, broadband HRRP data, and accompanying text description information of spatial micro-moving targets in the L-band, S-band, and X-band from radar systems in real time. The device first performs the same preprocessing (such as normalization and alignment) on the received multi-band multimodal data as in the training phase, and then inputs the preprocessed data into the deployed target recognition network. The feature extraction module in the target recognition network extracts the temporal fusion features, broadband HRRP features, and text features of the target RCS data in each frequency band. Subsequently, the multi-band graph attention module dynamically fuses these features from different frequency bands and modalities to obtain the final fused features. Finally, the classifier outputs the category label to which the spatial micro-moving target belongs. The device outputs the recognition results to a display terminal or subsequent decision-making system in real time, thereby completing the real-time recognition of spatial micro-moving targets.

[0034] Because the network fully integrates multi-band wide and narrow band radar data with text semantic information, and has a lightweight network structure (such as using efficient feature extractors such as 1-D TCN, HFE and Tiny-Bert), it can meet the real-time processing requirements while ensuring recognition accuracy.

[0035] To verify the superiority of this application, simulation experiments were conducted. Table 1 compares the recognition results of this application and existing technologies for spatial micro-moving targets at signal-to-noise ratios of 5dB, 10dB, 15dB, and 20dB. This application uses RCS data, HRRP data, and text data, while existing technologies use corresponding RCS data and HRRP data. Table 1 lists the results of this application and the recognition results of existing technologies. To facilitate the explanation of the superiority of this application in single-feature condition recognition compared to existing technologies, Table 1 also lists the recognition results under single-feature conditions.

[0036] Table 1 Comparison of noise reduction and recognition results between existing technologies and this application

[0037] The experimental results show that, at signal-to-noise ratios of 5dB, 10dB, 15dB, and 20dB, the recognition performance of this application is significantly improved compared to the prior art, demonstrating the superiority of this application.

[0038] It is worth noting that the terms "first" and "second" in this application are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0039] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of this application and should not be construed as limiting the specific implementation of this application to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of this application, and all such modifications or substitutions should be considered within the scope of protection of this application.

Claims

1. A spatial micro-movement target recognition method based on the fusion of multi-band data features and text information, characterized in that, include: S1: Divide the narrowband RCS data, wideband HRRP data and text data of spatial micro-movement targets in multiple frequency bands to generate training set and test set; S2: Construct a target recognition network that fuses multi-band data features and text information. The target recognition network includes a feature extraction module, a multi-band graph attention module, and a classifier. S3: Input the training set into the multi-band data feature and text information fusion recognition network, train it through the backpropagation algorithm, and then use the test set to test the trained target recognition network to obtain a target recognition network that meets the conditions. S4: Deploy the qualified target recognition network on the device to perform real-time recognition of spatial micro-moving targets.

2. The spatial micro-movement target recognition method based on multi-band data feature and text information fusion according to claim 1, characterized in that, S1 includes: Narrowband RCS data, broadband HRRP data, and text data of space micro-moving targets were acquired in L-band, S-band, and X-band. The initial elevation angle ranged from 31° to 43° in L-band, from 21° to 33° in S-band, and from 41° to 53° in X-band. The initial elevation angle varied uniformly at 1° intervals, the precession frequency varied uniformly at 0.2Hz intervals, and the precession angle varied uniformly at 0.15° intervals. 1573 sets of narrowband echo data were obtained for each type of target in each band, and Gaussian noise was added to obtain data with signal-to-noise ratios of 5dB, 10dB, 15dB, and 20dB. The samples corresponding to the initial pitch angles of 21° to 30°, 31° to 40°, and 41° to 50° were taken as the training set, and 4840 samples were obtained for the four types of targets in each frequency band; The samples corresponding to the initial pitch angles of 31° to 33°, 41° to 43°, and 51° to 53° were taken as the test set, and 1452 samples were obtained for the four types of targets in each frequency band.

3. The spatial micro-movement target recognition method based on multi-band data feature and text information fusion according to claim 1, characterized in that, The feature extraction module includes a target narrowband RCS feature extraction branch, a broadband HRRP feature extraction branch, and a text feature extraction branch; The target narrowband RCS feature extraction branch is constructed by sequentially connecting a 1-D TCN, a 1-D CNN, and a TFDF module. The 1-D TCN is used to extract the temporal features of the target narrowband RCS data, the 1-D CNN is used to extract the frequency domain features of the target narrowband RCS data, and the input of the TFDF module is connected to the output of the 1-D TCN and the output of the 1-D CNN, respectively, to fuse the temporal and frequency domain features and output the target narrowband RCS fused features. The wideband HRRP feature extraction branch is constructed by an HFE and is used to extract features from the wideband HRRP data and output wideband HRRP features. The text feature extraction branch is constructed by a pre-trained Tiny-Bert encoder and is used to encode text information and output text features. The input of the multi-band graph attention module is connected to the target narrowband RCS fusion feature, broadband HRRP feature and text feature output by the feature extraction module in each frequency band, respectively, and is used to dynamically fuse features of different frequency bands and different modalities to output the final fused feature. The input of the classifier is connected to the output of the multi-band graph attention module, and is used to output a multimodal classification result based on the final fused features.

4. The spatial micro-movement target recognition method based on multi-band data feature and text information fusion according to claim 3, characterized in that, The 1-D TCN includes a one-dimensional temporal convolutional layer, a weight normalization layer, a ReLU activation function, and a residual connection operation connected in sequence; wherein, the kernel size of the one-dimensional temporal convolutional layer is 2, the stride is 1, the dilation factor is 1, and the number of feature mapping channels is 16. The 1-D CNN includes a one-dimensional convolutional layer, a batch normalization layer, a Mish activation function, and a one-dimensional max pooling layer connected in sequence; wherein, the one-dimensional convolutional layer has a kernel size of 3, a stride of 1, a padding size of 1, and 16 feature mapping channels, and the one-dimensional max pooling layer has a kernel size of 2 and a stride of 2.

5. The spatial micro-movement target recognition method based on multi-band data feature and text information fusion according to claim 3, characterized in that, The TFDF module contains multiple time-domain and frequency-domain feature interaction layers, each composed of a TFC block; the TFC block processes time-domain and frequency-domain features in parallel, and the specific processing procedure is as follows: The input time-domain and frequency-domain features are split, and then linear transformations are performed on the discretized state transition matrix and projection matrix of the corresponding modes to obtain intermediate time-domain and intermediate frequency-domain features. The intermediate time-domain features and intermediate frequency-domain features are concatenated, then mapped to a specified dimension through a 1-D CNN, and then added to the input time-domain features and frequency-domain features respectively through residual connections to obtain the output time-domain features and output frequency-domain features of this layer. The output time-domain features and output frequency-domain features are used as inputs to the next layer TFC block.

6. The spatial micro-movement target recognition method based on multi-band data feature and text information fusion according to claim 5, characterized in that, The TFDF module also includes a pooling layer, a 1-D CNN, and a weighted layer connected in sequence; The input of the pooling layer is connected to the output of the last time-domain and frequency-domain feature interaction layer, and is used to perform max pooling operations on the final output time-domain features and frequency-domain features respectively to obtain pooled time-domain features and pooled frequency-domain features; the input of the one-dimensional convolutional neural network is connected to the output of the pooling layer, and is used to concatenate the pooled time-domain features and pooled frequency-domain features and map them to obtain dynamic weights; the input of the weighting layer is connected to the output of the pooling layer and the output of the one-dimensional convolutional neural network respectively, and is used to perform weighted summation on the pooled time-domain features and pooled frequency-domain features using the dynamic weights to output the target narrowband RCS fusion feature.

7. The spatial micro-movement target recognition method based on multi-band data feature and text information fusion according to claim 3, characterized in that, The broadband HFE comprises a linear layer, a convolutional layer, an activation function, a state-space model layer, and an MLP connected in sequence; the input HRRP sequence first undergoes a linear transformation through the linear layer, and then is processed by the convolutional layer and the activation function to obtain the initial features; The state space model layer is used to update the state and calculate the output of the initial features. Internally, it uses a discretized state space parameter matrix to recursively calculate the hidden state to obtain the state space output features. The MLP is used to perform nonlinear mapping on the state space output features to output high-resolution range image features.

8. The spatial micro-movement target recognition method based on multi-band data feature and text information fusion according to claim 3, characterized in that, The multi-band graph attention module is specifically used for: The target RCS fusion features, high-resolution range image features, and text features under multiple frequency bands are used as child nodes. The child nodes of different frequency bands are used to obtain multiple parent nodes through MLP, thereby constructing a graph structure. There are no edges connecting the child nodes, but there are edges connecting the parent nodes. The child nodes are only connected to the parent nodes of the corresponding frequency band. By propagating and aggregating the nodes in the graph structure through a graph attention network, multimodal fusion features under each frequency band are obtained; The multimodal fusion features under each frequency band are dynamically weighted to obtain the final fusion features.

9. The spatial micro-movement target recognition method based on multi-band data feature and text information fusion according to claim 1, characterized in that, The classifier uses a Softmax classifier, and its loss function is cross-entropy loss, which is calculated based on the one-hot encoding of the sample's true label and the probability output by the Softmax classification function.

10. The spatial micro-movement target recognition method based on multi-band data feature and text information fusion according to claim 1, characterized in that, Backpropagation is used when training the target recognition network, and the network parameters are adjusted in the direction of decreasing overall loss function. The overall loss function is the product of alignment loss and alignment loss coefficient, plus the product of recognition loss and recognition loss coefficient. The recognition loss is cross-entropy loss. The alignment loss is used to constrain the consistency of the distribution of different modal features in the embedding space.

Images