Ship target recognition method and device based on RCS sequence and electronic equipment

Abstract

The application provides a ship target identification method and device based on an RCS sequence and electronic equipment, a signal is transmitted through a transmitting device of a radar system, and a feedback signal is received through a receiving device; an improved OS-CNN model that is pre-trained is acquired; the feedback signal is input into the trained improved OS-CNN model, and the object type of the returned feedback signal is obtained. The improved OS-CNN network is used in the field of radar target identification, the classification recognition rate is high, and the application has practical popularization value and application value. By introducing the advantages of offset convolution, attention mechanism and full-scale convolution, the improved OS-CNN network model can effectively extract the characteristics of target and interference RCS sequence data while maintaining light weight and high efficiency, accurate classification is realized, and the improved low-resolution radar intelligent target identification classification capability has important practical significance.

Description

Technical Field

[0001] This invention belongs to the field of radar target recognition technology, specifically relating to a ship target recognition method, device, and electronic equipment based on RCS sequence. Background Technology

[0002] The ocean, vast and rich in resources, is an indispensable part of a nation, directly related to its sovereignty and territorial integrity. Continuous and thorough maritime surveillance is a fundamental support for a maritime power. Maritime targets are diverse, widely distributed, and their environments are unpredictable, making them difficult to identify and distinguish, and thus challenging to detect. Understanding the characteristics of maritime targets is crucial for effective detection. In actual maritime target detection, in addition to the target's own echo, various interference factors such as chaff, corner reflectors, and active external interference significantly impact identification performance in complex sea conditions. Therefore, improving radar's accurate identification of interference and targets has become a critical issue that urgently needs to be addressed. Radar corner reflectors are a typical passive jamming decoy in electronic countermeasures, possessing advantages such as low manufacturing cost and significant jamming effect. Corner reflectors strongly reflect incident radar waves along their initial path, resulting in a large scattering cross section, which can interfere with, deceive, and spoof radar systems. Analyzing the RCS (radar cross section) scattering characteristics of corner reflectors is of great significance, providing theoretical and data support for effectively identifying targets and corner decoys.

[0003] Radar is the primary means of detecting maritime targets. Radar target characteristics mainly include radar cross section (RCS), broadband characteristics, and polarization scattering moment. Currently, methods for identifying ships and corner reflectors include the Krogager polarization decomposition algorithm and methods based on micro-Doppler features. However, extracting the polarization information of targets or interference from radar is relatively difficult, increasing the complexity of the radar system and affecting the practicality of anti-jamming methods based on polarization processing. Strong clutter on the sea surface generates clutter spectra in the echo signal spectrum, also affecting the effectiveness of anti-jamming methods based on micro-Doppler features. RCS information, on the other hand, is narrowband information that almost all radars can obtain. RCS sequences are characterized by small data volume, good real-time performance, and relatively simple processing techniques. Compared to high-resolution one-dimensional range, RCS is easier to acquire and is an important basis for target selection strategies in conventional radar.

[0004] Existing radar target recognition technologies primarily rely on feature extraction based on human experience. This necessitates prior knowledge and results in low predictability and effectiveness. Furthermore, the algorithms are only applicable to specific situations and exhibit poor generalization. In contrast to traditional feature extraction, the development of deep learning technology offers a new direction for radar target recognition. Deep learning and neural networks are currently hot research topics in artificial intelligence, encompassing numerous theoretical methods. However, no single universal model can be applied to all data; even for the same problem, different parameters and hyperparameters can lead to significantly different results. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this invention provides a ship target identification method, apparatus, and electronic device based on RCS sequences. The technical problem to be solved by this invention is achieved through the following technical solution:

[0006] In a first aspect, the present invention provides a ship target identification method based on RCS sequences, comprising:

[0007] S100 transmits signals through the radar system's transmitter and receives feedback signals through the receiver.

[0008] The feedback signal includes the echo signal returned by the real ship target and the interference signal caused by the corner reflector;

[0009] S200, Obtain the pre-trained improved OS-CNN model; the improved OS-CNN model adopts two branch channels, and is implemented by adding a channel attention module and a feature fusion module to the OS-CNN model;

[0010] S300, the feedback signal is input into the trained improved OS-CNN model so that the improved OS-CNN model outputs the object type that returns the feedback signal.

[0011] Secondly, the present invention provides a ship target identification device based on RCS sequence, comprising:

[0012] The transmitter-receiver module transmits signals through the radar system's transmitter and receives feedback signals through the receiver.

[0013] The feedback signal includes the echo signal returned by the real ship target and the interference signal caused by the corner reflector;

[0014] The acquisition module is used to acquire a pre-trained improved OS-CNN model; the improved OS-CNN model is implemented by adding a channel attention module and a feature fusion module to the OS-CNN model;

[0015] A classification module is used to input the feedback signal into the trained improved OS-CNN model so that the improved OS-CNN model outputs the object type that returns the feedback signal.

[0016] Thirdly, the present invention provides an electronic device including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus;

[0017] Memory, used to store computer programs;

[0018] A processor, when executing a program stored in memory, implements the steps of the method described in any one of claims 1-8.

[0019] Beneficial effects:

[0020] This invention provides a method, apparatus, and electronic device for ship target recognition based on RCS sequences. The method involves transmitting signals through a radar system's transmitter and receiving feedback signals through a receiver. A pre-trained improved OS-CNN model is acquired. This improved OS-CNN model is implemented by adding a channel attention module and a feature fusion module to the existing OS-CNN model. The feedback signal is input into the trained improved OS-CNN model to obtain the object type returned in the feedback signal. This invention applies the improved OS-CNN network to radar target recognition, achieving a high classification accuracy and demonstrating practical application value. By incorporating the advantages of offset convolution, attention mechanisms, and full-scale convolution, the OS-CNN network model can effectively extract features from target and interference RCS sequence data while maintaining lightweight and high efficiency, achieving accurate classification with a classification accuracy of 95%. This has significant practical implications for improving the intelligent target recognition and classification capabilities of low-resolution radar.

[0021] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of a ship target identification method based on RCS sequence provided by the present invention;

[0023] Figure 2 This is a schematic diagram illustrating the process of training the improved OS-CNN model provided by the present invention;

[0024] Figure 3 This is a schematic diagram of the existing OS-CNN model structure provided by the present invention;

[0025] Figure 4 This is a schematic diagram of the improved OS-CNN model structure provided by the present invention;

[0026] Figures 5a-5b It is a schematic diagram of the simulated 3D model;

[0027] Figures 6a-6d These are schematic diagrams illustrating different sea state levels;

[0028] Figures 7-9 It shows the polarization patterns of three different objects: ships, corner reflectors, and corner reflector arrays.

[0029] Figures 10-11 This is a comparison of the RCS of different objects under vertical polarization conditions. Detailed Implementation

[0030] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0031] Before introducing the specific details of the present invention, the technical concept of the present invention will be introduced first.

[0032] This invention proposes to employ deep learning methods to study ship target recognition based on RCS sequences. A convolutional neural network model is constructed to achieve automatic feature extraction, which not only overcomes the dependence on feature extraction in traditional methods but also improves the accuracy of target recognition in cases of missing sample data. The matching research between deep neural networks and the ship target recognition problem, and the proposal of a deep neural network model suitable for distinguishing ships from interference, has significant practical implications.

[0033] The technical solution of the present invention will be described in detail below.

[0034] Combination Figures 1 to 4 This invention provides a ship target identification method based on RCS sequences, comprising:

[0035] S100 transmits signals through the radar system's transmitter and receives feedback signals through the receiver.

[0036] The feedback signal includes the echo signal returned by the real ship target and the interference signal caused by the corner reflector;

[0037] S200, Obtain the pre-trained improved OS-CNN model; the improved OS-CNN model is implemented by adding a channel attention module and a feature fusion module to the OS-CNN model;

[0038] S300, the feedback signal is input into the trained improved OS-CNN model so that the improved OS-CNN model outputs the object type that returns the feedback signal.

[0039] After the radar is powered on, it receives both echo signals from genuine ship targets and interference signals caused by corner reflectors. This invention, by understanding the radar target's RCS sequence, can utilize the differences between these two signals to distinguish between real and false targets, thereby correctly selecting the radar target for subsequent tracking and precision strikes. In the field of radar target recognition, after preprocessing the raw data, the design of the feature extractor determines the effectiveness of target recognition and is a key research area. Deep CNN models can achieve efficient feature extraction by building deep networks. Therefore, this invention aims to design a reasonable CNN architecture for RCS sequences, fully utilize the potential information of the RCS sequence, conduct matching research on deep neural networks and ship target recognition problems, and propose a deep neural network model suitable for identifying ships and interference. This has significant practical implications for improving the intelligent target recognition and classification capabilities of low-resolution radar.

[0040] In one specific embodiment of the present invention, reference is made to... Figures 2-4 The training process of the pre-trained improved OS-CNN model described in this invention includes:

[0041] S210, a static RCS database of pre-built ship, corner reflector and corner reflector array models;

[0042] Efficiently and accurately establishing ship and corner reflector models is fundamental for subsequent feature extraction and radar target identification. A theoretical derivation method was used to construct radar cross-section models of ships, corner reflectors, and their arrays. SolidWorks modeling and simulation software and 3D electromagnetic field simulation software were used to perform 3D modeling and electromagnetic simulation of the ships, corner reflectors, and their arrays to obtain a static RCS database, which serves as the target identification database for subsequent tasks.

[0043] S220, Construct an OS-CNN model, a channel attention module, and a feature fusion module, and modify the network structure of the OS-CNN model. Then, add the channel attention module and the feature fusion module to the modified OS-CNN model to obtain an improved OS-CNN model.

[0044] To address the problem of identifying and classifying interference from ships and corner reflectors, a target recognition model for ships and corner reflectors is constructed based on a deep neural network. (Reference) Figure 3 and Figure 4 This invention adds an attention layer to the OS-CNN (OMNI-SCALE CNN) network. While using the convolutional neural network to extract recognizable features from the RCS sequence, the attention mechanism guides the recognition model to focus on feature extraction of the target region during training, thereby improving the robustness and recognition performance of the model.

[0045] S230, the improved OS-CNN model is trained using the static RCS database to obtain the trained improved OS-CNN model.

[0046] In one specific embodiment of the present invention, S210 includes:

[0047] S211, obtain the dimensions of three objects: ship target, corner reflector, and corner reflector array;

[0048] S212, Using SolidWorks 3D modeling software, create 1:1 3D models of the three objects according to their dimensions;

[0049] refer to Figures 5a to 5b In practical applications, the most commonly used system is the icosahedral inflatable corner reflector system, which is composed of isosceles right triangles with a side length of 1 meter, and all three trihedral structures are identical in size and shape. The relevant performance parameters of the destroyer target are: length approximately 155 meters, beam approximately 20 meters, etc. Figure 3 As shown, a 1:1 3D model of the Arleigh Burke-class destroyer, corner reflectors, and their arrays was created using SolidWorks 3D modeling software, as follows. Figure 5a and Figure 5b As shown, it serves as the research object for subsequent RCS statistical feature extraction and target recognition.

[0050] S213, import the three-dimensional model into the three-dimensional electromagnetic simulation software to obtain the RCS amplitude characteristics of the ship target, corner reflector, corner reflector array and sea surface combination model within 360 degrees;

[0051] Combination Figures 6a-6d The present invention imports the established three-dimensional models into three-dimensional electromagnetic simulation software (computer simulation technology, CST) to obtain the RCS amplitude characteristics of the destroyer, the angular reflector and its array, and the combined sea surface model within 360 degrees, such as... Figure 7-9 As shown. The CST simulation parameters are set as follows: pitch angle θ∈[0, 90°] (step size 1°), frequency 5-6Hz, and algorithm is the Shooting and Bouncing Ray (SBR) method.

[0052] S214, set the frequency domain range and multiple frequency points within the frequency domain range, and obtain the RCS amplitude of the RCS amplitude characteristic at each frequency point;

[0053] This invention sets the frequency domain range to 5–6 GHz, forming an RCS frequency sequence with a bandwidth of 1 GHz, a step size of 200 MHz, and a total of 50 frequency points, wherein… Let be the target RCS amplitude obtained at the i-th frequency point.

[0054] S215, normalize the RCS amplitude and arrange the normalized data vertically to obtain a static RCS database.

[0055] To overcome the sensitivity of individual RCS amplitudes and prevent large amplitude values ​​from dominating, the data is normalized. This means discarding the true magnitude information and retaining only the shape and undulation information. The normalized data is denoted as... Perform maximum value normalization on the data.

[0056] refer to Figures 10 to 11 The RCS amplitude sequences of diagonal inverted targets, diagonal inverted arrays, and ship targets at 91*360 azimuth angles were processed as described above, and the final dataset with 361*90 rows * 51 columns was obtained by vertically arranging them as follows:

[0057]

[0058] Y = [C (1) C (2) ,…C (3) ,…C (32760) ];

[0059] In the above formula, X is the normalized data matrix; Y is the label vector; C∈[1,2,3] is the label category, 1 represents the inverted corner, 2 represents the inverted corner array, and 3 represents the ship.

[0060] In one specific embodiment of the present invention, reference is made to... Figure 4 The improved OS-CNN models in S220 include:

[0061] The system consists of two branch channels, a feature fusion module, and a fully connected layer. The first branch channel is composed of a channel attention module, and the second branch channel is composed of multiple custom convolutional layers, multiple channel attention modules, and a global average pooling layer. Each custom convolutional layer is followed by a channel attention module, and the last channel attention module is connected to the global average pooling layer. The first branch channel outputs a weighted artificial feature vector, and the second branch channel outputs a depth feature vector. The outputs of both branch channels are input to the feature fusion module, which performs feature fusion to obtain fused features. After passing through the fully connected layer, the system outputs the recognition and classification results.

[0062] refer to Figure 3(Custom Convolutional Layers) Multiple custom convolutional layers are constructed based on the parameter list configuration. Each convolutional layer has different parameters, so the shape of the input data changes with each layer's processing. The specific structure of the convolutional layer is implemented by the `nn.Conv1d` function, which is a one-dimensional convolutional layer used to process one-dimensional time-series data. In the constructor, a series of convolutional layers are generated based on the input parameter list. Each element in the parameter list represents the parameters of a convolutional layer, including the number of input channels, the number of output channels, the kernel size, and the stride. In the `forward` method, data is passed sequentially to each convolutional layer through a loop, and the ReLU activation function is applied. Then, a global average pooling layer is used to pool the features, flattening the pooled features, and then classification is performed through a fully connected layer.

[0063] refer to Figure 4 This invention adds an attention mechanism after each convolutional layer. The purpose of the attention mechanism is to enable the network to automatically focus on important parts of the input by learning attention weights. The attention mechanism uses a simple softmax function to calculate the attention weights and applies them to the feature tensor.

[0064] In one specific embodiment of the present invention, S230 includes:

[0065] S231, Training and testing sets are used separately for the static RCS database;

[0066] S232, Manual feature extraction is performed on the original data in the training set, and then the data is input into the first branch channel and the original data is input into the second branch channel;

[0067] S233, the attention weights are applied to the artificial feature tensor through the channel attention module of the first branch channel to obtain the weighted artificial feature vector and the depth feature vector is obtained through the second branch channel;

[0068] The first branch is the manual feature extraction branch plus the attention module (eca_layer). This invention performs manual feature extraction on the dataset, extracting a total of 28 features, including statistical and transform features. Then, channel attention enhancement is applied to the extracted features, resulting in weighted attention features with dimensions (batch_size, num_features = 28). The "eca_layer" module accepts input x, with dimensions (batch_size, channels, sequence_length), where channels are the number of channels in the input data, i.e., the number of features. Through the channel attention mechanism, the model can automatically learn the importance of feature channels, thus focusing more on features useful for the task. This helps improve model performance and generalization ability.

[0069] The second branch channel first uses adaptive average pooling to perform adaptive average pooling on the feature sequence output by the convolutional layer, reducing its dimension to a feature vector of (batch_size, num_features=448, 1). After removing the dimension of 1, a depth feature vector of (batch_size, num_features=448) is obtained.

[0070] S234, The artificial feature tensor and the deep feature vector are fused through the feature fusion module to obtain fused features;

[0071] S235 maps the fused features to the recognition and classification results through a fully connected layer;

[0072] S236, adjust the parameters of the improved OS-CNN model according to the loss function of the recognition and classification results, and return to S232 until the number of iterations is reached to obtain the improved OS-CNN model that has been trained.

[0073] Following S236, the ship target identification method based on RCS sequences further includes:

[0074] The improved OS-CNN model was tested using the test set.

[0075] This invention trains an improved OS-CNN model, receiving training and validation sets, along with corresponding label data. The training set consists of 15,000 samples. During training, an attention mechanism is used to enhance model performance. Model evaluation and results are printed every 50 iterations, for a total of 200 training iterations. First, the model is set to evaluation mode (eval()), and then the evaluation function is used to evaluate model performance on both the training and test sets, printing the training accuracy, test accuracy, and loss value. Next, the training results are saved to the results log folder, along with the model parameters.

[0076] This invention uses a pre-trained improved OS-CNN model for prediction. It receives test set data and labeled data, with the test set comprising 32,760 samples, and ultimately returns the model's prediction results and corresponding ground truth labels. The prediction accuracy is printed; using this model, the accuracy for classifying ships, corner reflectors, and corner reflector arrays is 95%.

[0077] In one specific embodiment of the present invention, the step of applying attention weights to the artificial feature tensor through the channel attention module of the first branch channel to obtain a weighted artificial feature vector includes:

[0078] a) The input features x extracted by artificial features are pooled by adaptive average pooling to obtain an artificial feature tensor y of shape (batch_size, channels, 1); in this way, the features of each channel are pooled into a scalar (1-dimensional).

[0079] b) Convolve the pooled artificial feature tensor y through a 1×1 convolutional layer to learn the attention weights for each channel; the artificial feature tensor y represents the attention weights for each channel; this tells the model which channels are more important for a specific task.

[0080] c. The Sigmoid activation function is used to restrict the output of the convolution to the range [0,1].

[0081] d. The input feature x is weighted and multiplied according to the channel attention weight y to obtain a weighted artificial feature vector. In this way, the feature of each channel will be strengthened or weakened according to the attention weight.

[0082] In one specific embodiment of the present invention, S300 includes:

[0083] S310, the feedback signal is manually feature extracted, and the manually extracted signal features are input into the first branch channel of the trained improved OS-CNN model and the feedback signal is input into the second branch channel;

[0084] S320, the signal features are processed through the first branch channel to obtain the weighted artificial feature vector of the feedback signal, and the feedback signal is processed through the second branch channel to obtain the depth feature vector of the feedback signal;

[0085] S330, The feature fusion module performs feature fusion on the depth feature vector and weighted artificial feature vector of the feedback signal to obtain the fused feature of the feedback signal;

[0086] S340, the fused features are classified through the fully connected classification layer to obtain the object type of the returned feedback signal.

[0087] This invention concatenates deep feature vectors and weighted artificial feature vectors through a fully connected layer, i.e., a feature fusion module, to obtain a feature vector of (batch_size, num_features = 448 + 28) = (batch_size, 476). Then, it uses softmax mapping to output the number of classes (batch_size, n_class = 3).

[0088] This invention achieves the following few-sample classification:

[0089] If few_shot = True (no few-shot classification):

[0090] The fused feature vectors are mapped through a fully connected layer to the output number of classes (batch_size, n_class = 3).

[0091] If few_shot = False (perform few-shot classification):

[0092] The output feature vector dimension is (batch_size, =448+1) = (batch_size, 449).

[0093] The final network output structure depends on the configuration of the convolutional layers in the parameter list. The attention module (eca_layer) enhances important information from the input features before concatenation. The few_shot function determines whether to perform few-shot classification. The final recognition and classification results are shown in Table 1.

[0094] Table 1 Final Classification Results

[0095]

[0096] As can be seen from Table 1, the classification accuracy of the improved OS-CNN model of this invention is 95%, which is a significant improvement compared to the existing accuracy.

[0097] This invention provides a ship target identification device based on RCS sequence, comprising:

[0098] The transmitter-receiver module transmits signals through the radar system's transmitter and receives feedback signals through the receiver.

[0099] The feedback signal includes the echo signal returned by the real ship target and the interference signal caused by the corner reflector;

[0100] The acquisition module is used to acquire a pre-trained improved OS-CNN model; the improved OS-CNN model is implemented by adding a channel attention module and a feature fusion module to the OS-CNN model;

[0101] A classification module is used to input the feedback signal into the trained improved OS-CNN model so that the improved OS-CNN model outputs the object type that returns the feedback signal.

[0102] This invention provides an electronic device including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus;

[0103] Memory, used to store computer programs;

[0104] The steps of a ship target identification method based on RCS sequence implemented by a processor when executing a program stored in memory.

[0105] This invention provides a method, apparatus, and electronic device for ship target recognition based on RCS sequences. The method involves transmitting signals through a radar system's transmitter and receiving feedback signals through a receiver. A pre-trained improved OS-CNN model is acquired. This improved OS-CNN model is implemented by adding a channel attention module and a feature fusion module to the existing OS-CNN model. The feedback signal is input into the trained improved OS-CNN model to obtain the object type returned in the feedback signal. This invention applies the improved OS-CNN network to radar target recognition, achieving a high classification accuracy and demonstrating practical application value. By incorporating the advantages of offset convolution, attention mechanisms, and full-scale convolution, the OS-CNN network model can effectively extract features from target and interference RCS sequence data while maintaining lightweight and high efficiency, achieving accurate classification with a classification accuracy of 95%. This has significant practical implications for improving the intelligent target recognition and classification capabilities of low-resolution radar.

[0106] Furthermore, the terms "first" and "second" 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. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0107] Although this application has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings, the disclosure, and the appended claims in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality.

[0108] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A ship target identification method based on RCS sequence, characterized in that, include: S100 transmits signals through the radar system's transmitter and receives feedback signals through the receiver. The feedback signal includes the echo signal returned by the real ship target and the interference signal caused by the corner reflector; S200, Obtain the pre-trained improved OS-CNN model; the improved OS-CNN model adopts two branch channels, and is implemented by adding a channel attention module and a feature fusion module to the OS-CNN model; S300, the feedback signal is input into the trained improved OS-CNN model so that the improved OS-CNN model outputs the object type that returns the feedback signal; The training process of the pre-trained improved OS-CNN model includes: S210, a static RCS database of pre-built ship, corner reflector and corner reflector array models; S220, Construct an OS-CNN model, a channel attention module, and a feature fusion module, and modify the network structure of the OS-CNN model. Then, add the channel attention module and the feature fusion module to the modified OS-CNN model to obtain an improved OS-CNN model. S230, the improved OS-CNN model is trained using the static RCS database to obtain the trained improved OS-CNN model; The improved OS-CNN models in S220 include: The system consists of two branch channels, a feature fusion module, and a fully connected layer. The first branch channel is composed of a channel attention module, and the second branch channel is composed of multiple custom convolutional layers, multiple channel attention modules, and a global average pooling layer. Each custom convolutional layer is followed by a channel attention module, and the last channel attention module is connected to the global average pooling layer. The first branch channel outputs a weighted artificial feature vector, and the second branch channel outputs a depth feature vector. The outputs of both branch channels are input to the feature fusion module, which performs feature fusion to obtain fused features. These features are then passed through a fully connected layer to output the recognition and classification results. S230 includes: S231, Training and testing sets are used separately for the static RCS database; S232, Manual feature extraction is performed on the original data in the training set, and then the data is input into the first branch channel and the original data is input into the second branch channel; S233, the attention weights are applied to the artificial feature tensor through the channel attention module of the first branch channel to obtain the weighted artificial feature vector and the depth feature vector is obtained through the second branch channel; S234, The artificial feature tensor and the deep feature vector are fused through the feature fusion module to obtain fused features; S235 maps the fused features to the recognition and classification results through a fully connected layer; S236, adjust the parameters of the improved OS-CNN model according to the loss function of the recognition and classification results, and return to S232 until the number of iterations is reached to obtain the improved OS-CNN model that has been trained.

2. The ship target identification method based on RCS sequence according to claim 1, characterized in that, S210 includes: S211, obtain the dimensions of three objects: ship target, corner reflector, and corner reflector array; S212, Using SolidWorks 3D modeling software, create 1:1 3D models of the three objects according to their dimensions; S213, import the three-dimensional model into the three-dimensional electromagnetic simulation software to obtain the RCS amplitude characteristics of the ship target, corner reflector, corner reflector array and sea surface combination model within 360 degrees; S214, set the frequency domain range and multiple frequency points within the frequency domain range, and obtain the RCS amplitude of the RCS amplitude characteristic at each frequency point; S215, normalize the RCS amplitude and arrange the normalized data vertically to obtain a static RCS database.

3. The ship target identification method based on RCS sequence according to claim 1, characterized in that, Following S236, the ship target identification method based on RCS sequences further includes: The improved OS-CNN model was tested using the test set.

4. The ship target identification method based on RCS sequence according to claim 1, characterized in that, The process of applying attention weights to the artificial feature tensor through the channel attention module of the first branch channel to obtain a weighted artificial feature vector includes: a. The input feature x extracted by artificial feature extraction is pooled by adaptive average pooling to obtain an artificial feature tensor y with shape (batch_size, channels, 1); b. Convolve the pooled artificial feature tensor y through a 1×1 convolutional layer to learn the attention weights for each channel; the artificial feature tensor y represents the attention weights for each channel. c. The Sigmoid activation function is used to restrict the output of the convolution to the range [0,1]. d, the input feature x is weighted and multiplied according to the channel attention weight y to obtain the weighted artificial feature vector.

5. The ship target identification method based on RCS sequence according to claim 1, characterized in that, The S300 includes: S310, the feedback signal is manually feature extracted, and the manually extracted signal features are input into the first branch channel of the trained improved OS-CNN model and the feedback signal is input into the second branch channel; S320, the signal features are processed through the first branch channel to obtain the weighted artificial feature vector of the feedback signal, and the feedback signal is processed through the second branch channel to obtain the depth feature vector of the feedback signal; S330, The feature fusion module performs feature fusion on the depth feature vector and weighted artificial feature vector of the feedback signal to obtain the fused feature of the feedback signal; S340, the fused features are classified through the fully connected classification layer to obtain the object type of the returned feedback signal.

6. A ship target identification device based on RCS sequence, characterized in that, include: The transmitter-receiver module transmits signals through the radar system's transmitter and receives feedback signals through the receiver. The feedback signal includes the echo signal returned by the real ship target and the interference signal caused by the corner reflector; The acquisition module is used to acquire a pre-trained improved OS-CNN model; the improved OS-CNN model is implemented by adding a channel attention module and a feature fusion module to the OS-CNN model; The classification module is used to input the feedback signal into the trained improved OS-CNN model so that the improved OS-CNN model outputs the object type that returns the feedback signal; The training process of the pre-trained improved OS-CNN model includes: S210, a static RCS database of pre-built ship, corner reflector and corner reflector array models; S220, Construct an OS-CNN model, a channel attention module, and a feature fusion module, and modify the network structure of the OS-CNN model. Then, add the channel attention module and the feature fusion module to the modified OS-CNN model to obtain an improved OS-CNN model. S230, the improved OS-CNN model is trained using the static RCS database to obtain the trained improved OS-CNN model; The improved OS-CNN models in S220 include: The system consists of two branch channels, a feature fusion module, and a fully connected layer. The first branch channel is composed of a channel attention module, and the second branch channel is composed of multiple custom convolutional layers, multiple channel attention modules, and a global average pooling layer. Each custom convolutional layer is followed by a channel attention module, and the last channel attention module is connected to the global average pooling layer. The first branch channel outputs a weighted artificial feature vector, and the second branch channel outputs a depth feature vector. The outputs of both branch channels are input to the feature fusion module, which performs feature fusion to obtain fused features. These features are then passed through a fully connected layer to output the recognition and classification results. S230 includes: S231, Training and testing sets are used separately for the static RCS database; S232, Manual feature extraction is performed on the original data in the training set, and then the data is input into the first branch channel and the original data is input into the second branch channel; S233, the attention weights are applied to the artificial feature tensor through the channel attention module of the first branch channel to obtain the weighted artificial feature vector and the depth feature vector is obtained through the second branch channel; S234, The artificial feature tensor and the deep feature vector are fused through the feature fusion module to obtain fused features; S235 maps the fused features to the recognition and classification results through a fully connected layer; S236, adjust the parameters of the improved OS-CNN model according to the loss function of the recognition and classification results, and return to S232 until the number of iterations is reached to obtain the improved OS-CNN model that has been trained.

7. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the steps of the method described in any one of claims 1-5.

Images