Sea surface target detection method based on data enhancement and recurrence plot classification of Transform technology

Through the data augmentation and recursive graph classification method based on Transformer technology, the problem of small target detection on sea surface is solved, the accuracy and efficiency of target detection on sea surface is improved, and the problem of complex sea conditions and data imbalance is adapted.

CN118980996AActive Publication Date: 2024-11-19NANJING UNIV OF POSTS & TELECOMM
View PDF 10 Cites 0 Cited by

Patent Information

Application Number
CN202411054612.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-02
Publication Date
2024-11-19
Estimated Expiration
2044-08-02

AI Technical Summary

Technical Problem

In the complex and changing sea conditions and data imbalance, it is difficult for the prior art to effectively detect and track small targets on the sea surface, especially in the case of strong sea clutter and rapid changes in sea conditions, the detection performance and robustness are limited.

Method used

Using data augmentation and recursive graph classification methods based on Transformer technology, the historical and future item data are generated through the preprocessing and bidirectional prediction mechanism of radar measured target echo amplitude data, and the recursive graph data set is constructed, and the classifier is trained using convolutional neural network. The model performs sea surface object detection.

Benefits of technology

It improves the accuracy and efficiency of target detection, can achieve excellent performance under unbalanced data proportions, and significantly improves the efficiency and classification effect of the processing process.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118980996A_ABST
    Figure CN118980996A_ABST
Patent Text Reader

Abstract

The invention discloses a sea surface target detection method based on data enhancement and recurrence plot classification of a Transform technology, and belongs to the field of radar target detection. The sea surface target detection method comprises the following steps: obtaining a sea surface echo amplitude sequence by using a radar, preprocessing the sea surface echo amplitude sequence to obtain radar actual measurement target echo amplitude data, and processing the radar actual measurement target echo amplitude data into a data format suitable for being sent to Transform; according to the preprocessed actually-measured target amplitude data, historical item data and future item data of the actually-measured target amplitude are generated by using a Transform bidirectional prediction mechanism, and the historical item data and the future item data are spliced with the actually-measured target amplitude data to obtain enhanced target amplitude data; constructing a recurrence plot data set by using the enhanced actually measured target amplitude data; training the recurrence plot data set by using a convolutional neural network to obtain a classifier model; and classifying the recurrence plot to be tested by using the classifier model, and outputting a classification result.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of radar target detection, and in particular relates to a sea surface target detection method based on data enhancement and recursive graph classification of Transformer technology. Background Art

[0002] Sea surface target detection is a key step in understanding marine scenes and is widely used in marine environmental monitoring, marine resource development, navigation safety and other fields. However, the complex statistical characteristics of sea clutter and its uneven distribution in space and time pose great challenges to target detection. Sea clutter is affected by many factors such as sea conditions, weather, radar parameters and operating frequency, and its intensity and distribution show a high degree of non-stationarity and non-Gaussian characteristics. Especially in strong clutter environments, the detection and tracking of small targets has become a global technical problem. Researchers have proposed a variety of detection methods to meet this challenge. Traditional methods mainly include time-frequency representation. Victor C. Chen et al. used information in the time domain and frequency domain to analyze and process signals in the article "Time-Frequency Transforms for Radar Imaging and Signal Analysis" to adapt to the changes in sea clutter. Deng Huawu et al. used specific filters to remove or suppress sea clutter and retain useful target signals in the article "Unsupervised Segmentation of Synthetic Aperture Radar Sea Ice Images Based on Markov Random Field Model". Although these traditional methods are effective under specific conditions, they are often difficult to adapt to complex and changeable sea conditions and data imbalance problems. In particular, their performance and robustness may be significantly limited in the case of strong sea clutter, rapidly changing sea conditions or scarce target samples. Summary of the invention

[0003] In view of the deficiencies in the prior art, the purpose of the present invention is to provide a sea surface target detection method based on data enhancement and recursive graph classification of Transformer technology, which solves the problems in the prior art.

[0004] The purpose of the present invention can be achieved through the following technical solutions:

[0005] The method for detecting sea surface targets based on data enhancement and recursive graph classification using Transformer technology is characterized by comprising the following steps:

[0006] The sea surface echo amplitude sequence is obtained by radar, and the radar measured target echo amplitude data is obtained by preprocessing, and then processed into a data format suitable for sending to Transformer;

[0007] According to the preprocessed measured target amplitude data, the Transformer bidirectional prediction mechanism is used to generate the historical item data and future item data of the measured target amplitude, and the historical item data and future item data of the measured target amplitude are spliced ​​with the measured target amplitude data to obtain the enhanced target amplitude data;

[0008] The enhanced measured target amplitude data are constructed into a recursive graph dataset;

[0009] The recursive graph dataset is trained using a convolutional neural network to obtain a classifier model; the classifier model is used to classify the recursive graph to be tested and the classification results are output.

[0010] The steps of preprocessing the sea surface echo amplitude sequence are:

[0011] S11, taking out the radar measured target data unit, adding its real part and imaginary part and then calculating the absolute value to obtain the radar measured target amplitude data;

[0012] S12, defining the radar measured target amplitude data as an item in the Transformer system, and setting a virtual user to interact with the item.

[0013] Furthermore, the sea surface echo amplitude sequence is composed of N consecutive pulse echoes received by the radar, forming a time series data set; each pulse echo data contains amplitude information.

[0014] Furthermore, the step of generating historical item data and future item data of the measured target amplitude includes:

[0015] S21, through reverse training from right to left, the original input data is reversed to generate historical input data and generate historical item data, the expression is:

[0016] p(v k-1 =v k |S)

[0017] In the formula, represents the original input sequence data set, v k Represents the first item of the original input sequence data; v k-1 Represents the first historical item data predicted based on the original input sequence data; this prediction process is repeated iteratively k times to obtain a historical item data set The superscript u indicates that the set is a set of sequence data interacting with user u;

[0018] S22, using As the input sequence, fine-tune the Transformer and predict the next item to generate future item data. The expression is:

[0019]

[0020] In the formula, It represents the enhanced data set obtained by splicing the predicted historical item data set and the original input sequence data set together in the order of the historical item data set and the original input sequence data set, that is, express The last term of Represents the first future item data predicted based on the enhanced data set; this prediction process is repeated iteratively k times to obtain the future item data set The historical item data set, the original input sequence data set and the future item data set are spliced ​​together to obtain the balanced target amplitude data set.

[0021] Furthermore, the step of constructing the recursive graph dataset includes:

[0022] S31, the equalized target amplitude data set is used as the input time series data to be divided into a number of sub-segments of length N, and for each data sub-segment x(t), an embedding vector set is constructed; then the embedding vector construction expression is:

[0023] v(t)={x(t),x(t+τ),x(t+2τ),...,x(t+(m-1)τ)}

[0024] Where v(t) is the vector generated by the embedding dimension and the embedding delay, i.e., the embedding vector, τ is the delay time used in the embedding process, and m is the embedding dimension;

[0025] S32, use the recursive matrix R to represent the distance between the embedded vectors; the element R of the recursive matrix R i,j The definition is as follows:

[0026] R i,j =||v(i)-v(j)||

[0027] Where 1≤i,j≤n, n=N-(m-1)τ, ||·|| represents the Manhattan distance; then the recursive matrix can be expressed as:

[0028]

[0029] S33, grayscale normalization processing is performed on the recursive matrix R to obtain a recursive graph data set; the formula for grayscale normalization processing is:

[0030]

[0031] Among them, G(i,j) is the grayscale value of the transformed image, R(i,j) represents the grayscale value of the original image; min and max represent the minimum and maximum grayscale values ​​in the original image respectively.

[0032] Furthermore, the specific steps of using the classifier model to classify the recursive graph to be tested include:

[0033] S41, pass the recursive graph dataset into the first convolutional layer of the convolutional neural network, and output the feature map of the recursive graph:

[0034]

[0035] Where f(·) is a nonlinear activation function; is the jth node of the convolutional layer; M i is the range of the local receptive field; k is the convolution kernel; b c is the bias; the output of the convolutional layer is the feature map of the recursive graph;

[0036] S42, input the feature map generated by the convolution layer into the next pooling layer:

[0037] z i =f down (y1,y2,...y i ),y i ∈V

[0038] Among them, f down (·) is the downsampling function. In the present invention, the maximum pooling method is adopted. i represents the output value of the i-th node, and V represents the value range of the pooling layer input value;

[0039] S43, the outputs of the convolutional layer and the pooling layer are input into the fully connected layer. The output of the fully connected layer is the classification result. The specific formula is as follows:

[0040]

[0041] in, Represents the output value of the i-th node, that is, the number of feature map elements, w and b f are the weight matrix and bias vector, f(·) represents the nonlinear activation function, represents the i-th node of the current fully connected layer, O f is the number of elements in the feature map output by the last layer.

[0042] The sea surface target detection system based on Transformer technology data enhancement and recursive graph classification includes:

[0043] Radar signal receiving module: used to obtain sea surface echo sequence;

[0044] Data preprocessing module: used to preprocess the sea surface echo data to obtain the radar measured sea surface echo amplitude data; according to the echo amplitude data obtained by preprocessing, it is processed again into a data format suitable for sending to Transformer;

[0045] Data generation module: Use the Transformer bidirectional prediction mechanism to generate historical item data and future item data of the measured target amplitude, and splice them with the measured target amplitude data to obtain enhanced target amplitude data;

[0046] Recursive graph structure generation module: constructs the enhanced measured target amplitude data into a recursive graph dataset;

[0047] Network model training module: used to train the recursive graph dataset and obtain the classifier model;

[0048] And, a classification module: used to classify the recursive graph to be tested using the trained classifier model and output the result.

[0049] A computer storage medium stores a readable program, which can execute the above-mentioned sea surface target detection method based on data enhancement and recursive graph classification based on Transformer technology when the program is run.

[0050] An electronic device, comprising: a processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other through the communication bus;

[0051] The memory is used to store at least one executable instruction, and the executable instruction enables the processor to perform operations corresponding to the above-mentioned sea surface target detection method based on data enhancement and recursive graph classification of Transformer technology.

[0052] A computer program product includes computer instructions, wherein the computer instructions instruct a computing device to perform operations corresponding to the above-mentioned sea surface target detection method based on data enhancement and recursive graph classification based on Transformer technology.

[0053] 1. The detection method of the present invention utilizes the data enhancement model of the Transformer network to generate high-quality historical and future data through a bidirectional prediction mechanism, and splices them to obtain enhanced data. The enhanced data is constructed into a recursive graph dataset, and the recursive graph, a tool for quantitative analysis of recursion and nonlinearity, is used to analyze sea clutter data, and a classifier model is obtained through training of a convolutional neural network. The classifier can obtain good classification results and improve the accuracy and effect of target detection.

[0054] 2. The present invention enhances data based on Transformer, which shows good performance compared with existing data enhancement technologies. Traditional methods usually require balanced data sets, such as a 1:1 ratio of target and non-target samples to obtain satisfactory results. The present invention proposes a data enhancement method based on Transformer technology, which can achieve excellent performance even under an unbalanced 3:10 ratio, improve efficiency, and speed up the entire processing process. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0056] Figure 1 is a flow chart of sea surface target detection of the present invention;

[0057] Figure 2 It is a flow chart of data enhancement based on Transformer technology of the present invention;

[0058] Figure 3 It is a performance comparison diagram under different sampling methods of the present invention;

[0059] Figure 4 It is a structural diagram of the sea surface target detection system of the present invention. DETAILED DESCRIPTION

[0060] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0061] like Figure 1 As shown in FIG. 1 , the sea surface target detection method based on data enhancement and recursive graph classification of Transformer technology includes the following steps:

[0062] S1, use radar to obtain the sea surface echo amplitude sequence, pre-process it to obtain the radar measured target echo amplitude data, and process it into a data format suitable for sending to Transformer.

[0063] The sea surface echo amplitude sequence is composed of N consecutive pulse echoes received by the radar, forming a time series data set. Each pulse echo data contains its corresponding amplitude information; the specific steps for preprocessing the sea surface echo amplitude sequence are as follows:

[0064] S11, taking out the radar measured target data unit, adding its real part and imaginary part and then calculating the absolute value to obtain the radar measured target amplitude data;

[0065] S12, defines the radar measured target amplitude data as "items" in the Transformer system, and sets virtual "users" to interact with these "items";

[0066] The number of projects contained in each user in the Transformer system is the number of interactions; there is a "timestamp" requirement in the Transformer system, and the timestamp is increased in sequence according to the arrangement order of the real radar data; other data format requirements in the Transformer system generate simulated data through random functions.

[0067] S2, based on the pre-processed measured target amplitude data, the Transformer bidirectional prediction mechanism is used to generate the historical item data and future item data of the measured target amplitude, and the historical item data, the measured target data, and the future item data are spliced ​​together to obtain the enhanced target amplitude data;

[0068] The specific steps to obtain the enhanced target amplitude data are:

[0069] S21, through reverse training from right to left, the original input data is reversed to generate possible historical input data, that is, historical item data similar to the real input data is generated. For a given sequence such as S = [v k ,v k+1 ,...,v n ],in Represents the kth to nth items in the original input data set S, all of which belong to the item set The formula for Transformer to predict historical item data is as follows:

[0070]

[0071] In the formula, Represents the original input sequence data set as v k Represents the first item of the original input sequence data; v k-1 Represents the first historical item data predicted based on the original input sequence data. This prediction process is repeated k times to obtain a historical item data set The superscript u indicates that the set is a set of sequence data interacting with user u.

[0072] This formula is explained as: by using the given original input data set S = [v k ,vk+1 ,...,v n ] predict v k-1 ,in Represents the previous item of the sequence. First, the sequence is input into the Transformer to perform an operation to generate a history item to obtain a history item v k-1 Then the history item v k-1 With the original sequence input v k to v n-1 Recombine as the input for the next operation. Repeat this process iteratively k-1 times, and finally generate k-1 All k historical items, i.e., the historical item data set represents the k-1th item interacted with user u.

[0073] Since the iterative prediction method is used, for the input sequence S = [v k ,v k+1 ,...,v n ], each iterative prediction will make these input sequences pass through the Transformer layer. This process promotes the Transformer network model to learn and integrate the contextual information of the previous elements, thereby producing richer and more advanced temporal dynamic representations and contextual dependencies. Specifically, for each sequence in the original input set S, the first prediction process generates a history item v k-1 At the same time, v k ,v k+1 ,...,v n After passing through the Transformer layer, it is converted into a new sequence representation that contains more comprehensive information. These new representations not only capture the internal relationship of the sequence, but also consider the interaction between items. Therefore, the representation of each element is based on the accumulated information of itself and all previous elements, so that the final historical item data set is Closer to the real data distribution. After multiple experimental verifications, the number of iterative predictions here is set to 52, the number of items in each input sequence set S is between 50 and 200, and the Transformer network model shows significant prediction performance. This parameter configuration can produce high-quality prediction results.

[0074] S22, using Fine-tune the Transformer as an input sequence and predict the next item;

[0075] Bundle As the input sequence, it is sent to the Transformer to fine-tune the parameters of the Transformer network model in detail so that the model can better understand and predict the dynamic changes of time series data. The principle is described in S21. After the fine-tuning is completed, it is used To make predictions, we get the prediction item v n+1 .

[0076] In order to continuously optimize the accuracy of prediction, the present invention adopts an iterative method for predicting future item data: after the first prediction is completed, The first term is truncated to obtain And v n+1 Fill to The last item gets It is then fed back into the Transformer network model for prediction, and this step is repeated k-1 times. Here, the formula used to predict the next item is:

[0077]

[0078] In the formula, It means that the predicted historical item data set and the original input sequence data set are concatenated in the order of the historical item data set and the original input sequence data set to obtain the enhanced data set. express The last term of Represents the first future item data predicted based on the enhanced data set. This prediction process is repeated k times, and we can get the future item data set We concatenate the historical item data set, the original input sequence data set, and the future item data set to obtain the balanced target amplitude data set. Right now Contains the historical data of the target unit predicted by Transformer, the measured target data, and the future data of the target unit predicted by Transformer.

[0079] S3, constructing the equalized target amplitude data set into a recursive graph data set;

[0080] The steps to construct a recursive graph dataset include:

[0081] S31, the enhanced measured target amplitude data set The input time series data is divided into several sub-segments of length N. For each data sub-segment x(t), an embedding vector set is constructed. The embedding vector construction expression is:

[0082] v(t)={x(t),x(t+τ),x(t+2τ),...,x(t+(m-1)τ)}

[0083] Where v(t) is the vector generated by the embedding dimension and the embedding delay, τ is the delay time used in the embedding process, and m is the embedding dimension;

[0084] S32, use the recursive matrix R to represent the distance between the embedded vectors; the element R of the recursive matrix R i,j The definition is as follows:

[0085] R i,j =||v(i)-v(j)||

[0086] Where 1≤i,j≤n, n=N-(m-1)τ, ||·|| represents the Manhattan distance; then the recursive matrix can be expressed as:

[0087]

[0088] S33, grayscale normalization processing is performed on the recursive matrix R to obtain a recursive graph data set; the formula for grayscale normalization processing is:

[0089]

[0090] Among them, G(i,j) is the gray value of the transformed image, R(i,j) represents the gray value of the original image; min and max represent the minimum and maximum gray values ​​in the original image respectively. Through the above method, the recursive graph dataset of the sea surface measured data is finally obtained.

[0091] S4, uses a convolutional neural network to train the recursive graph dataset, obtains a classifier model for classification, and finally obtains the classification result to realize the detection of sea surface targets.

[0092] The specific steps of using convolutional neural networks to train recursive graph datasets and obtain a classifier model include:

[0093] 1) First, build the network model and build the classic neural network model - LeNet-5, which includes convolution layer (C1), normalization layer (N2), activation layer (A3) pooling layer (C4), convolution layer (C5), normalization layer (N6), activation layer (A7), pooling layer (C8), fully connected layer (F9), and output layer (O10).

[0094] 2) Divide the obtained recursive graph dataset into two parts: a training set and a test set, where the ratio of the training set to the test set is 7:3.

[0095] 3) After data segmentation is completed, labels are added to the recursive graph to classify it into two categories: target and clutter. The training set and validation set are input into the network model, and after training, a classifier model is obtained.

[0096] The test set data is then input into the trained classifier model; the classifier model processes the input test set data based on the network parameters obtained through its training.

[0097] The classifier model is internally operated and quickly outputs the corresponding classification results; in this embodiment, the parameters for constructing the recursive graph are: time series length N = 1000, delay τ = 11, and embedding dimension m = 5. The network parameters are: the number of C1 convolution kernels is 16, the number of C3 convolution kernels is 32, and the convolution kernel size is 5x5. The activation layer uses the ReLU activation function. The pooling layer adopts the maximum pooling method, the pooling matrix size is 2x2, and the step size is 2.

[0098] The specific steps of using the trained convolutional neural network to classify recursive graph data include:

[0099] S41, the recursive graph dataset is passed into the first convolution layer of the convolutional neural network, and the feature map of the recursive graph is output. The specific formula is as follows:

[0100]

[0101] Where f(·) is a nonlinear activation function; is the jth node of the convolutional layer; M i is the range of the local receptive field; k is the convolution kernel; b c is the bias; the output of the convolutional layer is the feature map of the recursive graph;

[0102] S42, the feature map generated by the convolutional layer is input into the next pooling layer. The specific formula is as follows:

[0103] z i =f down (y1,y2,...y i ),y i ∈V

[0104] Among them, f down (·) is the downsampling function. In the present invention, the maximum pooling method is adopted. i represents the output value of the i-th node, and V represents the value range of the pooling layer input value.

[0105] S43, the output of the convolution layer and the pooling layer is input into the fully connected layer, and the output of the fully connected layer is the classification result. Specifically, the fully connected layer compresses the high-dimensional feature map into a low-dimensional representation. Each node in the last layer corresponds to the score or probability of a category. This process realizes the end-to-end mapping from the original recursive graph data to the final classification result. The specific formula is as follows:

[0106]

[0107] in, Represents the output value of the i-th node, that is, the number of feature map elements, w and b f are the weight matrix and bias vector, f(·) represents the nonlinear activation function, represents the i-th node of the current fully connected layer, O f is the number of elements in the feature map output by the last layer. Usually the first fully connected layer converts multiple feature maps into vectors, and the last fully connected layer acts as a softmax classification layer.

[0108] Example 2

[0109] In this embodiment, a specific experiment is conducted to verify the effect of the sea surface target detection method of the present invention.

[0110] In this embodiment, the data measured by Professor S. Haykin of McMaster University in Canada using IPIX radar in a real ocean environment is used as the radar to obtain the sea surface echo amplitude sequence of the present invention. In the experiment, the present invention method, simple oversampling, Smote oversampling, and random undersampling are used as comparison algorithms. The operation flow and some data of the present invention method are shown in FIG. Figure 2 As shown in Figure 2, the detection results of different algorithms under HH polarization are as follows: Figure 3 As shown in the figure, it can be seen that the three indicators of precision, recall and F1 coefficient of the method proposed in the present invention are all maintained at a high level, among which the most important F1 coefficient can reach 85.5%, while the F1 coefficients of simple oversampling, Smote oversampling and random undersampling methods are 79.8%, 80.54% and 73.01% respectively. Therefore, the data balancing method proposed in the present invention is superior to other classification methods in classification performance, improves efficiency and accelerates the entire processing process.

[0111] Example 3

[0112] like Figure 4 As shown, in this embodiment, a sea surface target detection system based on data enhancement and recursive graph classification of Transformer technology is proposed, including:

[0113] Radar signal receiving module: used to obtain sea surface echo sequence;

[0114] Data preprocessing module: used to preprocess the sea surface echo data to obtain the radar measured sea surface echo amplitude data; according to the echo amplitude data obtained by preprocessing, it is processed again into a data format suitable for sending to Transformer;

[0115] Data generation module: Use the Transformer bidirectional prediction mechanism to generate historical item data and future item data of the measured target amplitude, and splice the three together to obtain the enhanced target amplitude data;

[0116] Recursive graph structure generation module: receives the preprocessed recursive matrix as input and converts the matrix information into a data set of recursive graph structure through a specific algorithm;

[0117] Network model training module: used to train the recursive graph dataset and obtain the classifier model;

[0118] Classification module: used to classify the recursive graph to be tested using the trained classifier model and output the results.

[0119] The network model training module includes: a data set segmentation unit and a model training unit; wherein the data set segmentation unit is used to divide the obtained recursive graph into a training set and a validation set, and the ratio of the three is 7:3. The model training unit is used to use the obtained training set and validation set to perform classifier model training and parameter tuning verification, and finally obtain a classifier model.

[0120] The present invention enhances data based on Transformer, and shows good performance compared with existing data enhancement technologies. Traditional methods usually require balanced data sets, such as a 1:1 ratio of target and non-target samples to obtain satisfactory results. The present invention proposes a data enhancement method based on Transformer technology, which makes the classification results more accurate and reliable under an unbalanced 3:10 ratio, and the classification performance is more robust, which improves efficiency and speeds up the entire processing process.

[0121] The method of the present invention may be implemented in hardware, firmware, or as software or computer code that may be stored in a recording medium (such as a CDROM, RAM, floppy disk, hard disk, or magneto-optical disk), or as computer code that is originally stored in a remote recording medium or a non-temporary machine-readable medium downloaded over a network and will be stored in a local recording medium, so that the method described herein may be stored in such software processing on a recording medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware (such as an ASIC or FPGA). It is understood that a computer, processor, microprocessor controller, or programmable hardware includes a storage component (e.g., RAM, ROM, flash memory, etc.) that can store or receive software or computer code, and when the software or computer code is accessed and executed by a computer, processor, or hardware, the method described herein is implemented. In addition, when a general-purpose computer accesses the code for implementing the method shown herein, the execution of the code converts the general-purpose computer into a dedicated computer for executing the method shown herein.

[0122] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, and these changes and improvements all fall within the scope of the present invention to be protected.

Claims

1. A sea surface target detection method based on data enhancement and recursive graph classification using Transformer technology, characterized by: The following steps are involved: The radar is used to obtain the sea surface echo amplitude sequence, which is preprocessed to obtain the radar measured target echo amplitude data, and then processed into a data format suitable for input into the Transformer; Based on the pre-processed measured target amplitude data, the Transformer bidirectional prediction mechanism is used to generate the historical item data and future item data of the measured target amplitude, and the enhanced target amplitude data is obtained by splicing them with the measured target amplitude data. The enhanced measured target amplitude data are constructed into a recursive graph dataset; Use convolutional neural network to train the recursive graph dataset to obtain the classifier model; The classifier model is used to classify the recursive graph to be tested and the classification results are output.

2. The method for sea surface target detection based on data enhancement and recursive graph classification using Transformer technology according to claim 1, characterized in that: The steps of preprocessing the sea surface echo amplitude sequence are: S11, taking out the radar measured target data unit, adding its real and imaginary parts and then calculating the absolute value to obtain the radar measured target amplitude data; S12, defines the radar measured target amplitude data as an item in the Transformer system, and sets a virtual user to interact with the item.

3. The method for sea surface target detection based on data enhancement and recursive graph classification using Transformer technology according to claim 1 or 2, characterized in that: The sea surface echo amplitude sequence is composed of N consecutive pulse echoes received by the radar, forming a time series data set; each pulse echo data contains amplitude information.

4. The method for sea surface target detection based on data enhancement and recursive graph classification using Transformer technology according to claim 1, characterized in that: The steps of generating historical item data and future item data of the measured target amplitude include: S21, through reverse training from right to left, reversely infers the original input data to generate historical input data and generate historical item data, the expression is: Where, represents the original input sequence data set, v k Represents the first item of the original input sequence data; v k-1 Represents the first historical item data predicted based on the original input sequence data; this prediction process is repeated iteratively k times to obtain a set of historical item data The superscript u indicates that the set is a set of sequence data that interacts with user u; S22, using Fine-tune the Transformer as an input sequence and predict the next item to generate future item data. The expression is: Where, It represents the enhanced data set obtained by splicing the predicted historical item data set and the original input sequence data set together in the order of the historical item data set and the original input sequence data set, that is, express The last item of Represents the first future item data predicted based on the enhanced data set; this prediction process is repeated iteratively k times to obtain the future item data set The historical item data set, the original input sequence data set and the future item data set are spliced ​​together to obtain the balanced target amplitude data set.

5. The method for sea surface target detection based on data enhancement and recursive graph classification using Transformer technology according to claim 1, characterized in that: The steps of constructing the recursive graph dataset include: S31, the time series data of the target amplitude dataset after equalization is divided into several sub-segments of length N. For each data sub-segment x(t), an embedding vector set is constructed. The embedding vector construction expression is: v(t)={x(t),x(t+τ),x(t+2τ),...,x(t+(m-1)τ)} Where v(t) is the vector generated by embedding dimension and embedding delay, i.e., embedding vector, τ is the delay time used in the embedding process, and m is the embedding dimension; S32, use the recursive matrix R to represent the distance between the embedded vectors; the element R of the recursive matrix R i,j The definition is as follows: R i,j =||v(i)-v(j)|| Where 1≤i,j≤n, n=N-(m-1)τ, and ||·|| represents the Manhattan distance; the recursive matrix can be expressed as: S33, performing grayscale normalization processing on the recursive matrix R to obtain a recursive graph data set; the formula for grayscale normalization processing is: Among them, G(i,j) is the grayscale value of the transformed image, R(i,j) represents the grayscale value of the original image; min and max represent the minimum and maximum grayscale values ​​in the original image, respectively.

6. The method for sea surface target detection based on data enhancement and recursive graph classification using Transformer technology according to claim 1, characterized in that: The specific steps of using the classifier model to classify the recursive graph to be tested include: S41, passes the recursive graph dataset into the first convolutional layer of the convolutional neural network, and outputs the feature map of the recursive graph: Where f(·) is a nonlinear activation function; is the jth node of the convolutional layer; M i is the range of the local receptive field; k is the convolution kernel; b c is the bias; the output of the convolutional layer is the feature map of the recursive graph; S42, inputs the feature map generated by the convolutional layer into the next pooling layer: z i =f down (y1,y2,...y i ),y i ∈V Among them, f down (·) is the downsampling function. In this invention, the maximum pooling method is adopted. i Represents the output value of the i-th node, and V represents the range of the input value of the pooling layer; S43, the output of the convolutional layer and the pooling layer is input into the fully connected layer. The output of the fully connected layer is the classification result. The specific formula is as follows: in, Represents the output value of the i-th node, that is, the number of feature map elements, w and b f are the weight matrix and bias vector, f(·) represents the nonlinear activation function, represents the i-th node of the current fully connected layer, O f is the number of elements in the feature map output by the last layer.

7. A sea surface target detection system based on data enhancement and recursive graph classification using Transformer technology, characterized by: include: Radar signal receiving module: used to obtain sea surface echo sequence; Data preprocessing module: used to preprocess the sea surface echo data to obtain the radar measured sea surface echo amplitude data; Based on the echo amplitude data obtained by preprocessing, it is processed again into a data format suitable for sending to the Transformer; Data generation module: uses the Transformer bidirectional prediction mechanism to generate historical item data and future item data of the measured target amplitude, and splices them with the measured target amplitude data to obtain enhanced target amplitude data; Recursive graph structure generation module: constructs the enhanced measured target amplitude data into a recursive graph dataset; Network model training module: used to train the recursive graph dataset and obtain the classifier model; And, the classification module is used to classify the recursive graph to be tested using the trained classifier model and output the results.

8. A computer storage medium storing a readable program, characterized in that: When the program is running, it can execute the sea surface target detection method based on data enhancement and recursive graph classification based on Transformer technology described in any one of claims 1 to 6.

9. An electronic device, characterized in that: include: A processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other via the communication bus; The memory is used to store at least one executable instruction, and the executable instruction enables the processor to perform operations corresponding to the sea surface target detection method based on data enhancement and recursive graph classification based on Transformer technology as described in any one of claims 1-6.

10. A computer program product comprising computer instructions, characterized in that The computer instructions instruct the computing device to perform operations corresponding to the sea surface target detection method based on data enhancement and recursive graph classification based on Transformer technology as described in any one of claims 1-6.

Citation Information

Patent Citations

  • Time sequence classification method based on deep convolutional network and gray recursive graph analysis

    CN110443276A

  • Traffic data completion method and system, and storage medium

    CN113987310A

  • Trend prediction method based on attention mechanism and reinforcement learning

    CN114049222A

  • Sea surface target classification method and system

    CN114545354A

  • Audio signal processing method, and audio generation model training method and device

    CN114866856A