Radar target recognition method and device based on mobile window and attention model and medium

Abstract

The application discloses a radar target recognition method based on a moving window and an attention model, a computer device and a storage medium. The application trains a hierarchical shrinkage attention model based on a moving window, and uses the trained hierarchical shrinkage attention model for radar target recognition, so that the degree of manual participation in radar target recognition can be reduced, and the recognition accuracy can be improved. By using the trained hierarchical shrinkage attention model, a probability model for samples and a prior distribution do not need to be established, and only a single radar one-dimensional range image can be used for recognition, so that a large amount of pretreatment and feature extraction calculation of test data is not needed during recognition, and noise robustness is high. The application is easy to perform parallel calculation, so that the problem that a recurrent neural network is difficult to realize parallel calculation in the prior art can be overcome, and the problem that good recognition accuracy is obtained while the recognition efficiency is greatly improved can be solved. The application is widely applied to the technical field of radars.

Description

Technical Field

[0001] This invention relates to the field of radar technology, and in particular to a radar target recognition method, computer device, and storage medium based on a moving window and attention model. Background Technology

[0002] With the gradual development of radar technology, the bandwidth and resolution of modern radars have been continuously improved. The resolution of broadband radars is generally smaller than the size of the observed target. According to the scattering point model theory, the echo form of the target changes from a point target echo to the sum of echoes from the target's equivalent scattering center along the radar's line of sight, i.e., the sum of echoes from the target's important structural components. High-resolution one-dimensional range profiles (HRRPs) reflect the target's size and scattering center distribution, offering advantages such as strong real-time performance and ease of acquisition and storage, making them the primary means of identifying aerial targets such as aircraft. Therefore, quickly and accurately identifying targets using one-dimensional range profiles is crucial. However, high-resolution one-dimensional range profiles are easily affected by environmental noise, and in complex electromagnetic environments, various background noises and numerous interfering objects exist, impacting the performance of radar target identification methods.

[0003] Current technologies automatically extract separability features of one-dimensional range profiles of radar targets based on supervised learning methods, improving the shortcomings of traditional models in feature extraction. Deep learning-based radar high-resolution range profile recognition methods can be roughly divided into the following two categories: (1) Convolutional Neural Network (CNN) model; (2) Recurrent Neural Network (RNN) model. The former directly extracts and models features from the local envelope information of HRRP, and extracts and models features from the overall envelope through operations such as pooling, ignoring the sequence correlation that can reflect the physical structure features of the target. Although the latter models based on sequence correlation, it still has the following problems: (1) The original time-domain segmentation method is used for the local intensity information of HRRP, and the resulting features are highly redundant, which makes it difficult for subsequent RNN (Recurrent Neural Network) modeling; (2) When predicting, the unidirectional RNN can only use the structural information of the current time and the time before the current time, and cannot make good use of the overall structural prior information contained in HRRP; (3) RNN cannot operate in parallel and cannot utilize the parallel capabilities of hardware.

[0004] Terminology Explanation:

[0005] High-resolution one-dimensional range profile: A radar high-resolution range profile (HRRP) is the vector sum of the projections of the complex echoes from target scattering points acquired using broadband radar signals onto the radar ray. It provides information on the distribution of target scattering points along the range direction and is an important structural feature of the target. A radar high-resolution range profile can reflect the target's geometry and structural characteristics, and also contains more relevant information needed for target identification. Radar high-resolution range profiles can be used to measure parameters such as target height, velocity, and angle.

[0006] Deep learning originates from research on artificial neural networks. It discovers distributed feature representations of data by combining low-level features to form more abstract high-level representations of attribute categories or features.

[0007] Convolutional Neural Network: A convolutional neural network (CNN) is a deep learning model that uses convolution operations to extract features from data such as images. It is commonly used in fields such as image recognition and natural language processing.

[0008] Recurrent Neural Networks (RNNs): RNNs are a type of neural network architecture that can process sequential data, such as speech, text, and video. A key feature is that at each time step, in addition to input and output, there is a hidden state that remembers information from the previous time step and passes it to the next. There are various variants of RNNs, such as Long Short-Term Memory (LSTM) networks and Gated Recurrent Units (GRUs), which can solve the vanishing or exploding gradient problem and improve model performance. Summary of the Invention

[0009] In view of the technical problems in current radar technology, such as the failure to fully utilize the overall structural prior information contained in HRRP and poor parallel capabilities, the purpose of this invention is to provide a radar target recognition method, computer device and storage medium based on moving window and attention model.

[0010] On one hand, embodiments of the present invention include a radar target recognition method based on a moving window and an attention model, comprising:

[0011] Obtain a sample dataset; the sample dataset includes sample data of multiple radar targets;

[0012] A hierarchical shrinking attention model based on a moving window is established. This model includes a shrinking attention module, a discrete wavelet patch merging layer, and an output layer. The shrinking attention module comprises an embedding layer, a local perception module, a shrinking self-attention layer, a residual connection layer, and a convolutional feedforward layer. The embedding layer receives input data, performs patching and linear mapping on the input data, the local perception module extracts local features from the output data of the embedding layer, the shrinking self-attention layer calculates shrinking self-attention enhancement features within a window based on these local features, the residual connection layer and the convolutional feedforward layer sequentially perform residual connection and convolutional feedforward on the shrinking self-attention enhancement features, the discrete wavelet patch merging layer performs lossless downsampling on the output data of the shrinking attention module, and the output layer generates and outputs the classification result of the input data based on the output data of the discrete wavelet patch merging layer.

[0013] The hierarchical shrinking attention model is trained using the sample dataset.

[0014] The trained hierarchical contraction attention model is used for radar target recognition.

[0015] Furthermore, obtaining the sample dataset includes:

[0016] Create a CAD model of the radar target;

[0017] Using electromagnetic simulation software, the radar echo signal received is simulated based on the CAD model of the radar target.

[0018] The echo signal is converted into a high-resolution one-dimensional range image using a scattering point model.

[0019] The high-resolution one-dimensional distance image is used as the sample data to construct the sample dataset.

[0020] Furthermore, obtaining the sample dataset also includes:

[0021] Each of the aforementioned high-resolution one-dimensional range images is subjected to intensity normalization and sample alignment processing in sequence.

[0022] Further, the step of generating and outputting the classification result of the input data based on the output data of the discrete wavelet patch merging layer includes:

[0023] The output data of the discrete wavelet patch merging layer is subjected to layer normalization in the time dimension;

[0024] The result of layer normalization is then subjected to global average pooling.

[0025] The results of global average pooling are processed by full connection to obtain the normalized probabilities corresponding to each classification result.

[0026] Further, the step of calculating the contracted self-attention enhancement feature within the window for the local features includes:

[0027] Receive the local feature X output by the local perception module;

[0028] Through formula Perform layer normalization; where X norm For the layer normalization result, E(X) is the expected value of the local feature X, Var(X) is the variance of the local feature X, ε is a minimum number, γ is a learnable scaling parameter, and β is a learnable scalar parameter.

[0029] Based on the set window length, X norm Divide into multiple identical window blocks X win ;

[0030] For each of the aforementioned window blocks X win Perform relative position encoding to obtain the relative position encoding matrix B;

[0031] According to the formula Calculate the attention score matrix; where Z′ is the attention score matrix, and Q is the value for the window block X. win The query matrix obtained by performing a linear mapping, where K is the value for the window block X. win The key-value matrix obtained by performing a linear mapping, V is the value of the window block X. win The value matrix obtained by linear mapping, V′ is the result of applying shrinkage threshold feature enhancement to the value matrix V;

[0032] The attention score matrix Z′ is fully connected to obtain the contracted self-attention enhancement feature Z; the contracted self-attention enhancement feature Z is used to perform residual connection and convolutional feedforward on the residual connection layer and the convolutional feedforward layer to obtain the output data output of the contracted attention module.

[0033] Furthermore, the lossless downsampling includes:

[0034] The output data of the contraction attention module is subjected to discrete wavelet transform to obtain a first frequency component and a second frequency component; the frequency range corresponding to the first frequency component is higher than the frequency range corresponding to the second frequency component.

[0035] The first frequency component and the second frequency component are concatenated in the channel dimension to obtain the output data output′ of the discrete wavelet patch merging layer.

[0036] Further, training the hierarchical shrinking attention model using the sample dataset includes:

[0037] Perform several training cycles; in each training cycle, perform the following steps:

[0038] Obtain sample data from the sample dataset;

[0039] The sample data is input into the hierarchical contraction attention model for processing to obtain the category information prediction value of the hierarchical contraction attention model;

[0040] Obtain the true value of the category information corresponding to the sample data;

[0041] The loss function is determined based on the predicted value of the category information and the actual value of the category information;

[0042] Based on the loss function, the parameters of the hierarchical shrinking attention model are updated through backpropagation.

[0043] When the loss obtained by validating the hierarchical contractile attention model converges, the training of the hierarchical contractile attention model ends; otherwise, the next training cycle is executed.

[0044] Furthermore, training the hierarchical shrinking attention model using the sample dataset further includes:

[0045] When the training loop number is even, before inputting the sample data into the hierarchical shrinking attention model, the sample data is cyclically shifted using half the window length as the shift length.

[0046] On the other hand, embodiments of the present invention also include a computer device, including a memory and a processor, the memory being used to store at least one program, and the processor being used to load the at least one program to execute a radar target recognition method based on a moving window and attention model in the embodiments.

[0047] On the other hand, embodiments of the present invention also include a storage medium storing a processor-executable program, which, when executed by a processor, is used to perform a radar target recognition method based on a moving window and attention model in the embodiments.

[0048] The beneficial effects of this invention are as follows: The radar target recognition method based on moving window and attention model in the embodiments reduces the degree of human intervention in radar target recognition and improves the recognition accuracy by training a hierarchical shrinking attention model based on moving window and using the trained hierarchical shrinking attention model for radar target recognition. By using the trained hierarchical shrinking attention model, there is no need to establish a probability model for the samples or introduce prior distributions. Recognition can be performed using only a single one-dimensional radar range image. During recognition, there is no need to perform a large amount of preprocessing and feature extraction calculations on the test data, and it is highly robust to noise. The hierarchical shrinking attention model in this embodiment is implemented based on the principles of shrinking attention module, moving window, and hierarchy. When implementing the hierarchical shrinking attention model by computer, it is easy to perform parallel computing, thereby overcoming the problem that it is difficult to achieve parallel computing when using recurrent neural networks in current related technologies. Parallel computing can easily improve recognition efficiency while obtaining good recognition accuracy. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating the steps of the radar target recognition method based on a moving window and attention model in the embodiment.

[0050] Figure 2 This is a schematic diagram of the hierarchical shrinking attention model based on a moving window in the embodiment.

[0051] Figure 3 This is a schematic diagram illustrating the principle of the local sensing module in the embodiment;

[0052] Figure 4 This is a schematic diagram illustrating the principle of shrinking the self-attention layer in the embodiment;

[0053] Figure 5 This is a schematic diagram illustrating the principle of the shrinkage threshold feature enhancement module in the embodiment;

[0054] Figure 6 This is a schematic diagram illustrating the principle of the discrete wavelet patch merging layer in the embodiment;

[0055] Figure 7 This is a schematic diagram of the output layer in the embodiment. Detailed Implementation

[0056] In this embodiment, refer to Figure 1 A radar target recognition method based on a moving window and attention model includes the following steps:

[0057] S1. Obtain the sample dataset;

[0058] S2. Establish a hierarchical shrinking attention model based on a moving window;

[0059] S3. Train the hierarchical shrinking attention model using the sample dataset;

[0060] S4. The trained hierarchical shrinkage attention model is used for radar target recognition.

[0061] When performing step S1, which is to obtain the sample dataset, the following steps can be performed:

[0062] S101. Set the CAD model of the radar target;

[0063] S102. Run electromagnetic simulation software to simulate the echo signal received by the radar based on the CAD model of the radar target;

[0064] S103. Using a scattering point model, the echo signal is converted into a high-resolution one-dimensional range image;

[0065] S104. Construct a sample dataset using high-resolution one-dimensional distance images as sample data.

[0066] In step S101, the CAD model of the radar target can correspond to the same aircraft model.

[0067] In step S103, the principle of the scattering point model is as follows: the resolution of a broadband radar is generally smaller than the size of the observed target. Thus, the echo generated by the target illuminating the broadband radar is transformed from a point target echo into the sum of echoes from the target's equivalent scattering center along the radar's line-of-sight. In other words, by combining the echo signals representing important structural components of the target, a high-resolution one-dimensional range profile (HRRP) can be obtained. The high-resolution one-dimensional range profile reflects the target's size and the distribution of its scattering centers.

[0068] In step S104, a sample dataset is constructed using high-resolution one-dimensional distance images as sample data. This sample dataset can be divided into a training dataset, a validation dataset, and a test dataset in a 1:1:1 ratio, used for training, validating, and testing the hierarchical shrinking attention model, respectively. For example, the training dataset selected from the sample dataset is denoted as T = {(x...} i ,y i )} i∈[1,n],k∈[1,c] , where x i Let y represent the i-th sample. k This indicates that the sample belongs to the k-th class, and a total of c classes of targets were collected, where n represents the total number of samples.

[0069] After performing step S1 and before performing step S2, preprocessing can be performed on each sample data in the sample dataset, i.e., each high-resolution one-dimensional range image. In this embodiment, the preprocessing of the high-resolution one-dimensional range image includes the following steps:

[0070] P1. Intensity normalization, assuming the original HRRP is represented as x raw = [x1,x2,…,x L ], where L represents the total number of distance cells contained in the HRRP, then the intensity-normalized HRRP is expressed as:

[0071]

[0072] P2. Sample alignment: Translate the HRRP so that its centroid g is near L / 2. This way, the distance cells containing information in the HRRP will be distributed near the center. The calculation method for the HRRP centroid g is as follows:

[0073]

[0074] Where, x i After intensity normalization and centroid alignment, the amplitude of the original HRRP sample, which is the i-th distance unit in the original HRRP, has been limited to between 0 and 1. This not only unifies the scale, but also makes the values ​​between 0 and 1 very beneficial for subsequent neural network processing. HRRP echo signals that are distributed to the right or left have been adjusted to be near the center point.

[0075] By executing steps P1 and P2, the original HRRP echo is processed using L2 intensity normalization before HRRP processing, thereby improving the intensity sensitivity of HRRP. Since HRRP is extracted from radar echo data through a range window, the position of the range image acquired during the extraction process is not fixed in the range gate, resulting in translation sensitivity of HRRP. Steps P1 and P2 use centroid alignment to eliminate translation sensitivity, so that the range cells containing information in HRRP will be distributed near the center, which can make training and testing have a unified standard.

[0076] In step S2, the structure of the established hierarchical shrinking attention model based on the moving window is as follows: Figure 2 As shown.

[0077] Reference Figure 2 The hierarchical shrinking attention model based on moving windows includes a shrinking attention module, a discrete wavelet patch merging layer, and an output layer. The shrinking attention module comprises an embedding layer, a local perception module, a shrinking self-attention layer, a residual connection layer, and a convolutional feedforward layer.

[0078] The embedding layer receives input data, performs patching and linear mapping on the input data, the local perception module extracts local features from the output data of the embedding layer, the shrinking self-attention layer calculates shrinking self-attention enhanced features within a window for the local features, the residual connection layer and the convolutional feedforward layer perform residual connection and convolutional feedforward sequentially on the shrinking self-attention enhanced features, the discrete wavelet patch merging layer performs lossless downsampling on the output data of the shrinking attention module, and the output layer generates and outputs the classification result of the input data based on the output data of the discrete wavelet patch merging layer.

[0079] Specifically, when training the hierarchical shrinking attention model based on a moving window, the input data for the hierarchical shrinking attention model based on a moving window is the sample data in the sample dataset in step S1, i.e., HRRP; when using the hierarchical shrinking attention model based on a moving window for radar target recognition, the input data for the hierarchical shrinking attention model based on a moving window is data of the same type as the sample data, which is also HRRP.

[0080] When the HRRP is input into a hierarchical shrinking attention model based on a moving window, the embedding layer receives the input data, namely the HRRP. The embedding layer performs patching and linear mapping on the HRRP. Specifically, the embedding layer encodes the HRRP using convolution to obtain L segments of feature encoding, each segment having a dimension of N. The L segments of feature encoding are then output to the local perception module.

[0081] In this embodiment, the principle of the local sensing module is as follows: Figure 3 As shown. (Refer to...) Figure 3 The local perception module consists of residual depthwise convolutions. The formula for the local perception module is as follows:

[0082]

[0083] Among them, F out F represents the feature map output by the local perception module. in The input features are the output data of the embedding layer. represents the convolution operation, and K represents the depthwise convolution kernel.

[0084] In this embodiment, the principle of shrinking the self-attention layer is as follows: Figure 4 As shown. (Refer to...) Figure 4 The self-attention layer receives local features X (X has the shape [batch size, N, L], where batch size is the number of items in each batch, L is the length of X, and N is the dimension at each time point) from the local perception module, and then uses the formula... Perform layer normalization. Where X normFor the layer normalization result, E(X) is the expected value of the local feature X, Var(X) is the variance of the local feature X, ε is a minimum number, γ is a learnable scaling parameter, and β is a learnable scalar parameter.

[0085] Next, based on the set window length (e.g., W), X... norm Divide into multiple identical window blocks X win Each window block X win The shape is [batch size×(L / W),N,W].

[0086] Next, shrink the self-attention layer for each window block X. win Relative position encoding is performed to obtain the relative position encoding matrix B. In this embodiment, relative position encoding is an encoding method used to represent the relative distance and direction between different time points within a window. It can enhance the spatial perception capability of the hierarchical shrinking attention model of the moving window, enabling it to distinguish time points at different locations. When performing relative position encoding, firstly, two indices are assigned to each time point within the window. One index represents the relative distance in the x-direction, and the other index represents the relative distance in the y-direction. For example, if the window size is 1*7, then each pixel has an x-index and a y-index, where x is 0 and y ranges from -6 to 6. Then, the corresponding embedding vector is found in two predefined embedding matrices, and they are concatenated to form the relative position encoding. This results in a 7*7*d-dimensional relative position encoding matrix, where d is the embedding dimension.

[0087] After obtaining the relative position encoding matrix B, shrink the self-attention layer to divide N window blocks X win The data is divided into k groups, and then these k groups are linearly mapped to obtain the query matrix Q, the key matrix K, and the value matrix V. The value matrix V is then input into the shrinkage threshold feature enhancement module for processing to obtain V′.

[0088] In this embodiment, the principle of the shrinkage threshold feature enhancement module is as follows: Figure 5 As shown. (Refer to...) Figure 5 The shrinkage threshold feature enhancement module can execute the formula y = sign(x) × max(|x| - τ, 0) for calculation. Here, y is the output, x is the input, and τ is the learned threshold, satisfying... If the absolute value of the input signal is less than the threshold, the output is 0; if the absolute value of the input signal is greater than the threshold, the output is the input minus the threshold multiplied by the sign of the result. This allows the threshold to be dynamically adjusted according to the characteristics of the input V. If a part of the input V is important, its corresponding weight will be close to 1, resulting in a larger threshold; if a part of the input signal is not important or is noise, its corresponding weight will be close to 0, resulting in a smaller threshold. This effectively enhances the effective information in the feature map and suppresses noise and useless information.

[0089] In this embodiment, the formula The principle is as follows: α is a fixed positive number obtained by taking the absolute value of V and then averaging it channel by channel. σ is the Sigmoid activation function, W and b are the parameters of the one-dimensional convolution, and V is the input signal. First, the input V is passed through a one-dimensional convolutional layer to obtain a new feature vector. Then, the Sigmoid activation function is applied to this feature vector, transforming it into values ​​between 0 and 1. These values ​​are the weights of the threshold, reflecting the importance of different parts of the input V. Finally, these weights are multiplied by a fixed positive number α to obtain the final threshold.

[0090] Shrink self-attention layer execution formula Calculate the attention score matrix Z′, and then input Z′ into the fully connected layer to obtain the contracted self-attention enhancement feature Z.

[0091] In this embodiment, the shrinking self-attention enhancement feature Z is first normalized by the residual connection layer to realize the residual connection processing of the shrinking self-attention enhancement feature Z. This can effectively eliminate the information loss problem caused by increasing the number of layers, quickly backpropagate the loss, avoid gradient vanishing, and is beneficial for stacking deep neural networks.

[0092] Next, the normalized result of the residual connection layer is input into the convolutional feedforward layer. The convolutional feedforward layer consists of two convolutional layers and a Gaussian error linear unit (GELU) activation function. The first convolutional layer maps the N-dimensional features to a higher dimension and then adds non-linearity through the GELU activation function. The second convolutional layer maps from the high dimension to the low dimension, further preserving useful information.

[0093] The GELU activation function takes the form GELU(x) = x × P(X ≤ x) = x × Φ(x), x ~ N(0,1), where x is the input value, X is a random variable following a standard normal distribution, and Φ(x) is the cumulative distribution function of the standard normal distribution. A characteristic of this activation function is that the smaller the input x is, the smaller the probability of it being multiplied by a 0 or 1, which is equivalent to performing random dropout on the input.

[0094] The data output by the convolutional feedforward layer is used as the output data of the contraction attention module, and is denoted as output.

[0095] In this embodiment, the principle of the discrete wavelet patch merging layer is as follows: Figure 6 As shown. (Refer to...) Figure 6 In the Discrete Wavelet Patch Merging Layer, when performing lossless downsampling on the output data of the contraction attention module, a discrete wavelet transform is applied to the output data to obtain a first frequency component and a second frequency component. The frequency range corresponding to the first frequency component is higher than that corresponding to the second frequency component; therefore, the first frequency component can also be called a high-frequency component, and the second frequency component can also be called a low-frequency component. The Discrete Wavelet Patch Merging Layer then concatenates the first and second frequency components along the channel dimension to obtain the output data output′ of the Discrete Wavelet Patch Merging Layer.

[0096] In this embodiment, if the shape of the output is [batch_size, L, N], then the shape of output' is [batch_size, L / 2, 2N], thus achieving lossless downsampling of the output. After obtaining output', it can be passed through a convolutional layer for information exchange, allowing for full interaction between high-frequency and low-frequency components. Residual structures can also be used to avoid information loss, enabling rapid backpropagation of the loss and preventing gradient vanishing.

[0097] In this embodiment, the principle of the output layer is as follows: Figure 7 As shown. (Refer to...) Figure 7 The output layer performs layer normalization on the output data (output′) of the discrete wavelet patch merging layer in the time dimension. Then, the result of the layer normalization is processed by a global average pooling layer to reduce dimensionality, thus greatly reducing the network parameters. Regularization of the entire algorithm network structure prevents overfitting. Finally, a fully connected layer processes the result of the global average pooling to obtain the normalized probability corresponding to each classification result.

[0098] Specifically, the output layer can use the Softmax function to predict the normalized probability of HRRP for each aircraft category. The Softmax function is defined as follows: In the formula, the symbol z i Let z represent the output value of the i-th node, and let C represent the number of output nodes, i.e. the number of categories. Softmax(z) represents the output value of multi-class classification. One-hot encoding can be used to encode the HRRP dataset. One-hot encoding solves the problem that classifiers cannot handle attribute data well and also plays a role in expanding features to a certain extent.

[0099] Based on the working principle of the hierarchical shrinking attention model based on the moving window, if HRRP type data is input into the hierarchical shrinking attention model based on the moving window, the hierarchical shrinking attention model based on the moving window will output normalized probability type data to represent the probability of the aircraft type corresponding to the HRRP identified by the hierarchical shrinking attention model, that is, the identification result of the target corresponding to the HRRP.

[0100] Based on the working principle of the hierarchical shrinking attention model based on the moving window described above, the following steps can be performed when executing step S3, which is to train the hierarchical shrinking attention model using the sample dataset:

[0101] Perform several training cycles; in each training cycle, perform the following steps:

[0102] S301. Obtain sample data from the sample dataset;

[0103] S302. Input the sample data into the hierarchical contraction attention model for processing, and obtain the category information prediction value of the hierarchical contraction attention model;

[0104] S303. Obtain the true value of the category information corresponding to the sample data;

[0105] S304. Determine the loss function based on the predicted values ​​and the actual values ​​of the category information;

[0106] S305. Update the parameters of the hierarchical shrinking attention model through backpropagation based on the loss function;

[0107] S306. When the loss obtained by validating the hierarchical contraction attention model converges, the training of the hierarchical contraction attention model ends; otherwise, the next training cycle is executed.

[0108] Each training cycle includes steps S301-S306. Taking one training cycle as an example, steps S301-S306 will be explained.

[0109] In step S301, a sample data, namely an HRRP, is read from the sample dataset (specifically, the training dataset).

[0110] In step S302, the sample data is input into the hierarchical shrinking attention model. The embedding layer, shrinking attention module, discrete wavelet patch merging layer and output layer in the hierarchical shrinking attention model are processed in sequence to output the normalized probability corresponding to each classification result. The classification result with the largest normalized probability can be selected as the predicted value of the category information.

[0111] In step S303, the true value of the category information corresponding to the sample data is obtained, that is, the information of the true category of the target corresponding to the HRRP input to the hierarchical shrinking attention model (e.g., the aircraft type corresponding to HRRP).

[0112] In step S304, the degree of inconsistency between the predicted value and the true value of the category information is calculated using the cross-entropy loss function. In this embodiment, the loss function is defined as follows: Where p = [p0,...,p C-1 ] is a probability distribution, where each element p i This represents the probability that a sample belongs to the i-th class; [y = [y0,...,y...] C-1 ] is the One-hot representation of the Iwamoto label; when the sample belongs to the i-th class, y i =1, otherwise y i =0; c is the sample label.

[0113] In step S305, during the backpropagation phase, the AdamW optimizer is used to update the weight parameters in the hierarchical shrinking attention model based on the value of the loss function obtained in step S304.

[0114] In step S306, sample data (HRRP type data) is read from the validation set and input into the hierarchical contraction attention model after the weight parameters are updated in step S305. The loss function between the output of the hierarchical contraction attention model and the true value of the category information corresponding to the sample data is obtained. When the value of the loss function converges (for example, the value of the loss function itself or its fluctuation is less than a certain threshold), the training of the hierarchical contraction attention model ends, that is, the next training cycle is not executed. Otherwise, the next training cycle is executed.

[0115] In this embodiment, the executed training loops can be counted, and the sequence number of each training loop can be recorded, i.e., which training loop is being executed. When the sequence number of the executed training loop is even (e.g., the 2nd, 4th, etc. training loop), before inputting the sample data into the hierarchical shrinking attention model, the sample data can be cyclically shifted using half the window length W as the shift length before being input into the hierarchical shrinking attention model for training.

[0116] By executing steps S301-S306 several times, the hierarchical shrinking attention model based on the moving window is trained.

[0117] By executing steps S1-S3, a trained hierarchical contraction attention model based on a moving window can be obtained. This hierarchical contraction attention model has the ability to process HRRP type data and identify the target type (e.g., aircraft type) corresponding to the HRRP type data.

[0118] In step S4, the HRRP to be identified is input into the trained hierarchical shrinking attention model based on moving windows. The identification result of the HRRP can be determined based on the output of the hierarchical shrinking attention model (e.g., the various target types that the HRRP may correspond to and their probabilities, and the target type with the highest probability can be determined as the final identification result).

[0119] In this embodiment, the principle of executing steps S1-S4 is as follows: by setting a local perception module in the hierarchical shrinking attention model, the problem that the conventional transformer model cannot effectively extract local features can be solved; the shrinking self-attention layer in the hierarchical shrinking attention model can realize the shrinking self-attention mechanism, thereby enhancing the effective signal in the feature map, suppressing noise and useless signals, and enhancing the feature expression capability of the transformer model; the discrete wavelet patch merging layer in the hierarchical shrinking attention model can reduce the spatial resolution of the input sequence while increasing the number of channels, thereby allowing more efficient computation and information flow in the subsequent layers of the network.

[0120] Based on the principles of steps S1-S4, the radar target recognition method based on moving window and attention model in this embodiment can achieve the following technical effects:

[0121] By training a hierarchical contraction attention model based on a moving window, and then using the trained hierarchical contraction attention model for radar target recognition, the degree of human intervention in radar target recognition can be reduced, thereby improving recognition accuracy. Using the trained hierarchical contraction attention model eliminates the need to build probability models for samples or introduce prior distributions; recognition can be performed using only a single one-dimensional radar range image. During recognition, extensive preprocessing and feature extraction calculations of the test data are not required, and the model exhibits strong robustness to noise. The hierarchical contraction attention model in this embodiment is implemented based on principles such as contraction attention modules, moving windows, and hierarchy. When implementing the hierarchical contraction attention model using a computer, parallel computation is easily performed, overcoming the difficulty of achieving parallel computation in current related technologies using recurrent neural networks. Parallel computation can significantly improve recognition efficiency while maintaining good recognition accuracy.

[0122] A computer program can be written to execute a radar target recognition method based on a moving window and attention model as described in this embodiment. This computer program can be written into a computer device or storage medium. When the computer program is read out and run, the radar target recognition method based on a moving window and attention model as described in this embodiment can be executed, thereby achieving the same technical effect as the radar target recognition method based on a moving window and attention model as described in this embodiment.

[0123] It should be noted that, unless otherwise specified, when a feature is referred to as "fixed" or "connected" to another feature, it can be directly fixed or connected to the other feature, or indirectly fixed or connected to the other feature. Furthermore, the descriptions of "upper," "lower," "left," and "right" used in this disclosure are only relative to the relative positional relationships of the various components of this disclosure in the accompanying drawings. The singular forms "a," "described," and "the" used in this disclosure are also intended to include the plural forms, unless the context clearly indicates otherwise. Moreover, unless otherwise defined, all technical and scientific terms used in this embodiment have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this embodiment specification is only for describing particular embodiments and is not intended to limit the invention. The term "and / or" as used in this embodiment includes any combination of one or more of the associated listed items.

[0124] It should be understood that although the terms first, second, third, etc., may be used to describe various elements in this disclosure, these elements should not be limited to these terms. These terms are only used to distinguish elements of the same type from each other. For example, a first element may also be referred to as a second element without departing from the scope of this disclosure, and similarly, a second element may also be referred to as a first element. The use of any and all instances or exemplary language (“e.g.,” “such as,” etc.) provided in this embodiment is intended only to better illustrate embodiments of the invention and, unless otherwise required, does not impose a limitation on the scope of the invention.

[0125] It should be recognized that embodiments of the present invention can be implemented or carried out by computer hardware, a combination of hardware and software, or by computer instructions stored in a non-transitory computer-readable storage medium. The method can be implemented using standard programming techniques—including a non-transitory computer-readable storage medium configured with a computer program, wherein such a storage medium causes the computer to operate in a specific and predefined manner—according to the methods and drawings described in the specific embodiments. Each program can be implemented in a high-level procedural or object-oriented programming language to communicate with the computer system. However, if desired, the program can be implemented in assembly or machine language. In any case, the language can be a compiled or interpreted language. Furthermore, for this purpose, the program can run on a programmed application-specific integrated circuit (ASIC).

[0126] Furthermore, the procedures described in this embodiment can be performed in any suitable order unless otherwise indicated by this embodiment or clearly contradicted by the context. The procedures (or variations and / or combinations thereof) described in this embodiment can be executed under the control of one or more computer systems configured with executable instructions, and can be implemented by hardware or a combination thereof as code (e.g., executable instructions, one or more computer programs, or one or more applications) that commonly executes on one or more processors. The computer program includes a plurality of instructions executable by one or more processors.

[0127] Furthermore, the method can be implemented in any suitable type of computing platform, including but not limited to personal computers, minicomputers, mainframes, workstations, networked or distributed computing environments, standalone or integrated computer platforms, or in communication with charged particle tools or other imaging devices. Aspects of the invention can be implemented as machine-readable code stored on a non-transitory storage medium or device, whether removable or integrated into a computing platform, such as a hard disk, optical read and / or write storage medium, RAM, ROM, etc., such that it is readable by a programmable computer, and when the storage medium or device is read by the computer, it can be used to configure and operate the computer to perform the processes described herein. Furthermore, the machine-readable code, or portions thereof, can be transmitted via wired or wireless networks. The invention described in this embodiment includes these and other different types of non-transitory computer-readable storage media when such media comprises instructions or programs that implement the steps described above in conjunction with a microprocessor or other data processor. When programmed according to the methods and techniques described in the invention, the invention also includes the computer itself.

[0128] A computer program can be applied to input data to perform the functions described in this embodiment, thereby transforming the input data to generate output data stored in non-volatile memory. The output information can also be applied to one or more output devices, such as a display. In a preferred embodiment of the invention, the transformed data represents physical and tangible objects, including specific visual depictions of physical and tangible objects generated on the display.

[0129] The above description is merely a preferred embodiment of the present invention. The present invention is not limited to the above-described embodiments. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention, as long as they achieve the technical effects of the present invention by the same means, should be included within the scope of protection of the present invention. Within the scope of protection of the present invention, the technical solutions and / or implementation methods can have various modifications and variations.

Claims

1. A radar target recognition method based on a moving window and attention model, characterized in that, The radar target recognition method based on moving window and attention model includes: Obtain a sample dataset; the sample dataset includes sample data of multiple radar targets; A hierarchical shrinking attention model based on a moving window is established. This model includes a shrinking attention module, a discrete wavelet patch merging layer, and an output layer. The shrinking attention module comprises an embedding layer, a local perception module, a shrinking self-attention layer, a residual connection layer, and a convolutional feedforward layer. The embedding layer receives input data, performs patching and linear mapping on the input data, the local perception module extracts local features from the output data of the embedding layer, the shrinking self-attention layer calculates shrinking self-attention enhancement features within a window based on these local features, the residual connection layer and the convolutional feedforward layer sequentially perform residual connection and convolutional feedforward on the shrinking self-attention enhancement features, the discrete wavelet patch merging layer performs lossless downsampling on the output data of the shrinking attention module, and the output layer generates and outputs the classification result of the input data based on the output data of the discrete wavelet patch merging layer. The hierarchical shrinking attention model is trained using the sample dataset. The trained hierarchical contraction attention model is used for radar target recognition.

2. The radar target recognition method based on a moving window and attention model according to claim 1, characterized in that, The acquisition of the sample dataset includes: Create a CAD model of the radar target; Using electromagnetic simulation software, the radar echo signal received is simulated based on the CAD model of the radar target. The echo signal is converted into a high-resolution one-dimensional range image using a scattering point model. The high-resolution one-dimensional distance image is used as the sample data to construct the sample dataset.

3. The radar target recognition method based on a moving window and attention model according to claim 2, characterized in that, The process of obtaining the sample dataset also includes: Each of the aforementioned high-resolution one-dimensional range images is subjected to intensity normalization and sample alignment processing in sequence.

4. The radar target recognition method based on a moving window and attention model according to claim 1, characterized in that, The step of generating and outputting the classification result of the input data based on the output data of the discrete wavelet patch merging layer includes: The output data of the discrete wavelet patch merging layer is subjected to layer normalization in the time dimension; The result of layer normalization is then subjected to global average pooling. The results of global average pooling are processed by full connection to obtain the normalized probabilities corresponding to each classification result.

5. The radar target recognition method based on a moving window and attention model according to claim 1, characterized in that, The calculation of contracted self-attention enhancement features within a window for the local features includes: Receive the local features output by the local perception module ; Through formula Perform layer normalization; where, For the layer normalization result, For the local features Expected value For the local features variance It is a very small number. For learnable scaling parameters, Inexpensive parameters that can be learned; Based on the set window length, Divide into multiple identical window blocks ; For each of the aforementioned window blocks Perform relative position encoding to obtain the relative position encoding matrix. ; According to the formula Calculate the attention score matrix; where, The attention score matrix is... For the window block The query matrix obtained by performing a linear mapping. For the window block The key-value matrix obtained by performing a linear mapping, For the window block The value matrix obtained by performing a linear mapping, For the value matrix The result of performing shrinkage threshold feature enhancement; For the attention score matrix Perform a fully connected operation to obtain the contracted self-attention enhancement feature. The contraction self-attention enhancement feature This is used to perform residual connections and convolutional feedforwards between the residual connection layer and the convolutional feedforward layer to obtain the output data of the contraction attention module. .

6. The radar target recognition method based on a moving window and attention model according to claim 5, characterized in that, The lossless downsampling includes: The output data of the contraction attention module Perform discrete wavelet transform to obtain a first frequency component and a second frequency component; the frequency range corresponding to the first frequency component is higher than the frequency range corresponding to the second frequency component. The first frequency component and the second frequency component are concatenated in the channel dimension to obtain the output data of the discrete wavelet patch merging layer. .

7. The radar target recognition method based on a moving window and attention model according to any one of claims 1-6, characterized in that, The step of training the hierarchical shrinking attention model using the sample dataset includes: Perform several training cycles; in each training cycle, perform the following steps: Obtain sample data from the sample dataset; The sample data is input into the hierarchical contraction attention model for processing to obtain the category information prediction value of the hierarchical contraction attention model; Obtain the true value of the category information corresponding to the sample data; The loss function is determined based on the predicted value of the category information and the actual value of the category information; Based on the loss function, the parameters of the hierarchical shrinking attention model are updated through backpropagation. When the loss obtained by validating the hierarchical contractile attention model converges, the training of the hierarchical contractile attention model ends; otherwise, the next training cycle is executed.

8. The radar target recognition method based on a moving window and attention model according to claim 7, characterized in that, The step of training the hierarchical shrinking attention model using the sample dataset further includes: When the sequence number of the training loop being executed is even, before inputting the sample data into the hierarchical shrinking attention model, the sample data is cyclically shifted using half the window length as the shift length.

9. A computer device, characterized in that, The system includes a memory and a processor, the memory being used to store at least one program, and the processor being used to load the at least one program to execute the radar target recognition method based on a moving window and attention model as described in any one of claims 1-8.

10. A computer-readable storage medium storing a processor-executable program, characterized in that, The processor-executable program, when executed by the processor, is used to perform the radar target recognition method based on a moving window and attention model as described in any one of claims 1-8.

Images