Radar target signal identification method and system based on self-attention mechanism

Abstract

The invention provides a radar target signal identification method and system based on a self-attention mechanism, and the method comprises the steps: enabling a radar original echo signal to be divided through a sliding window operation, enabling the divided signal to serve as a model input, constructing a convolutional neural network in which features are extracted in parallel through a plurality of receptive field channels, fully exploring the multi-level information in data, and achieving the recognition of a radar target signal. The method achieves the comprehensive extraction and mutual complementation of features, remarkably improves the detection precision, introduces a new loss function, and effectively solves the problem of target position mismatching. The loss function enables the model to adaptively adjust the difference between the target and the background during training, thereby improving the learning ability and detection performance of the model. The method shows performance advantages under different signal-to-noise ratio conditions, the improvement amplitude is unequal from 2% to 25%, and the method is particularly remarkable in a low-signal-to-noise-ratio environment. The performance of the method in a target detection task is superior to that of a traditional method, and the method can be effectively applied to real-time detection of small targets on the sea and has good application prospects.

Description

Technical Field

[0001] The present invention relates to the technical field of radar signal processing, and particularly to a radar target signal recognition method and system based on a self-attention mechanism. Background Art

[0002] Sea target detection is one of the most important tasks in the field of sea radar and has also been a research hotspot in recent years. This task focuses on whether there is a target in each range cell of the radar echo in a specific sea area. However, in actual scenarios, due to the presence of severe sea clutter, noise, and human interference on the sea surface, small targets on the sea surface are usually buried in various interferences and are difficult to directly detect. In traditional sea radar target detection algorithms, CFAR (Constant False Alarm Rate) is one of the most widely used algorithms, and methods such as CA-CFAR (Cell Averaging CFAR) and OS-CFAR (Order Statistic CFAR) have also been developed from the traditional CFAR algorithm. However, when applying the above algorithms in actual scenarios, since the echo energy of small sea targets is often less than the energy of surrounding clutter, a large number of false alarms and missed alarms occur.

[0003] Convolutional neural networks (CNNs) have shown excellent performance in fields such as computer vision due to their good feature extraction ability. For example, in a published paper (Sun Ningyuan, Chen Xiaolong, etc., "Intelligent Detection Method for Radar Sea Targets Based on Dual-Channel Convolutional Neural Network Feature Fusion", "Modern Radar", Issue 10, 2019), an intelligent detection method for radar sea targets based on a dual-channel convolutional neural network (DCCNN) was proposed. In addition, in a published paper (Xing Hongyan, Zhu Qingqing, Xu Wei, etc., "A Method for Detecting Weak Signals under Chaotic Sea Clutter Background", "Acta Physica Sinica", Vol. 63, No. 10, 2014), a nonlinear system based on the phase space reconstruction theory was proposed, and the genetic algorithm prediction method was combined with the support vector machine for research. Currently, most CNN-based sea target detection algorithms usually process the echo after pulse compression into a time-frequency diagram or an RP diagram and then perform classification. However, the format of such radar images loses the inherent information of the signal, that is, key information such as phase. And in methods such as directly performing STFT on the original data to obtain a time-frequency diagram or performing data insertion to generate an RP diagram, when detecting each range cell, directly classifying the time-frequency diagram or RP diagram cannot obtain the target position.

[0004] When applying CNN to one-dimensional sequence processing, such as using one-dimensional convolutional kernels, it can be used for the classification of sequence samples. In a published paper (Hu Tao et al., "Research on Convolutional Neural Networks in Abnormal Sound Recognition", "Signal Processing", Vol. 34, No. 3, 2018), a one-dimensional convolutional kernel CNN network was used to process abnormal sound recognition, and the problems of signal anti-noise performance and abnormal signal processing of the CNN network when processing one-dimensional signals were analyzed. Another example is that in a published paper (Yin Heyi, Guo Zunhua et al., "One-Dimensional Convolutional Neural Networks for Radar High-Resolution Range Profile Recognition", "Telecommunication Engineering", Vol. 58, No. 10, 2018), one-dimensional CNN was used for the recognition of radar high-resolution range profile signals, and it was also verified that one-dimensional CNN signals have a good recognition effect on target signals in radar. However, one-dimensional detection methods lack surrounding environment or context information, which limits the understanding of the relationship between the target and the background by target detection algorithms. Summary of the Invention

[0005] Object of the Invention: The first object of the present invention is to provide a radar target signal recognition method based on self-attention mechanism that can retain the information of radar original echo signals and accurately extract them, and the second object is to provide a radar target signal recognition system corresponding to the above method.

[0006] Technical Solution: A radar target signal recognition method based on self-attention mechanism includes the following steps:

[0007] S1. Collect radar original echo signals and construct an original data matrix using the radar original echo signals;

[0008] S2. Perform a sliding window operation on the original data matrix obtained in step S1, intercept the radar original echo signals into several sub-data matrices of a set size, label the target signals and sea clutter signals in the radar original echo signals, accumulate radar echo data, and divide the obtained sub-data matrices into a training set, a validation set, and a test set according to a set ratio;

[0009] S3. Construct a convolutional neural network, which includes a Focus layer, a first receptive field channel and a second receptive field channel respectively connected to the Focus layer, and a fully connected layer respectively connected to the first receptive field channel and the second receptive field channel, and use the two receptive field channels to extract features in parallel;

[0010] S4. Input the training set into the convolutional neural network obtained in step S3, update the weights of the convolutional neural network using the gradient descent method, minimize the loss function and update the parameters to obtain a trained convolutional neural network;

[0011] S5. Input the validation set into the convolutional neural network trained in step S4, set the false alarm rate of the convolutional neural network, and obtain a set threshold through the validation set to obtain a radar target signal recognition model. Input the test set into the radar target signal recognition model, perform target detection on the test set, and output the recognition accuracy of the test set.

[0012] Specifically, in step S1, the data matrix is a discrete sequence with the distance cells at fast time on the horizontal axis and the number of pulses at slow time on the vertical axis.

[0013] Specifically, in step S2, perform a sliding window operation on the data matrix obtained in step S1, and intercept the original radar echo signal into several data matrices of size 64*128, where 64 is the distance cell at fast time and 128 is the accumulated number of pulses at slow time.

[0014] Specifically, in step S2, the annotation of the target signal and sea clutter signal in the original radar echo signal includes: marking the position of the distance cell where the target is located in the data matrix as 1, and the position of the distance cell where the sea clutter is located as 0.

[0015] Specifically, in step S3, both the first receptive field channel and the second receptive field channel include:

[0016] The first convolutional layer, the first dimension transformation layer, the multi-head self-attention module, the data stacking dimension transformation layer, the second convolutional layer, the residual block, the second dimension transformation layer, the third convolutional layer, and the third dimension transformation layer connected in sequence.

[0017] Specifically, the convolution kernel size of the first convolutional layer of the first receptive field channel is 3×3, the stride is 1×1, and the padding is (1, 1); the convolution kernel size of the first convolutional layer of the second receptive field channel is 3×8, the stride is 1×3, and the padding is (1, 0).

[0018] Specifically, in step S4, the expression of the loss function is:

[0019] Loss total = λLoss ce +(1 - λ)Loss mse

[0020] In the formula: Loss ce is the classification loss for evaluating the presence or absence of a target, Loss mse is the regression loss taking the prediction of the target position as, and λ is the weight factor for balancing the regression and classification losses.

[0021] Specifically, in the loss function, the calculation formula of the classification loss Loss ce is:

[0022]

[0023] Where: N is the number of cells, yi is the label of the i-th range cell, pi is the predicted value of the i-th range cell, and σ(p i ) is the probability that the target exists in the i-th range cell.

[0024] Specifically, in the loss function, the regression loss Loss mse is calculated as follows:

[0025]

[0026] Where: y i t is the probability that the target exists at the target position, p i t is the label, and n is the number of targets.

[0027] The present invention also provides a radar target signal recognition system based on a self-attention mechanism, including:

[0028] Data matrix construction module: used to collect the original radar echo signal and construct an original data matrix using the original radar echo signal;

[0029] Data partitioning module: used to perform a sliding window operation on the original data matrix, intercept the original radar echo signal into several sub-data matrices of a set size, label the target signal and sea clutter signal in the original radar echo signal, accumulate the radar echo data, and divide the obtained sub-data matrices into a training set, a validation set, and a test set according to a set ratio;

[0030] Model construction module: used to construct a convolutional neural network, the convolutional neural network includes a Focus layer, a first receptive field channel and a second receptive field channel respectively connected to the Focus layer, and a fully connected layer respectively connected to the first receptive field channel and the second receptive field channel, and uses the two receptive field channels to extract features in parallel;

[0031] Model training module: used to input the training set into the convolutional neural network, update the weights of the convolutional upgrade network using the gradient descent method, minimize the loss function and update the parameters to obtain the trained convolutional neural network;

[0032] Model validation module: used to input the validation set into the trained convolutional neural network, set the false alarm rate of the convolutional neural network, and obtain a set threshold through the validation set to obtain a radar target signal recognition model, input the test set into the radar target signal recognition model, perform target detection on the test set, and output the recognition accuracy of the test set.

[0033] Beneficial effects: Compared with the prior art, the remarkable effect of the present invention is:

[0034] 1. The present invention uses the original radar echo signal as the input of the model for direct multi-target detection on the sea surface. Different from the prior art that requires a large amount of preprocessing (such as STFT, RP maps, etc.) of the original data, it directly processes the signal using the designed complex self-attention mechanism model. The preprocessing in the prior art can also be regarded as a combination and replacement of convolution calculations. Since the convolutional neural network has a powerful fitting ability, it can replace the traditional processing methods and quickly obtain the calculation results.

[0035] 2. The convolutional neural network designed by the present invention can extract information related to echo pulses in different range cells from multiple receptive fields. By using multiple receptive fields to extract features of the same input data, although the input data is the same, the features extracted by different receptive fields are different. By integrating multiple receptive fields, the extracted features can be complemented, thereby greatly improving the detection performance of the convolutional neural network, and then realizing the rapid detection of radar target signals, and finally realizing the function of real-time target detection.

[0036] 3. The present invention designs a new loss function for model training, which solves the problems of target localization in the original echo, the imbalance in the number of targets and sea clutter, and the uneven distribution of targets from the perspective of the loss function. The problem of small target detection on the sea surface is transformed into a binary classification problem for each cell. This loss function has two detection heads, that is, it uses a binary classification head and a regression head to process data, and can accurately locate the position of the detected target. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 is the flowchart of the method of the present invention.

[0038] Figure 2 is the structural diagram of the convolutional neural network of the present invention.

[0039] Figure 3 is the echo amplitude diagram of small targets on the sea surface of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0040] The following further describes the solution of the present invention with reference to the accompanying drawings.

[0041] Embodiment 1

[0042] Please refer to Figure 1 shown. This embodiment provides a method for identifying radar target signals based on the self-attention mechanism, including the following steps:

[0043] S1. Collect the original radar echo signal, and further accumulate the signal to obtain a discrete data matrix S with the horizontal axis being the range cell and the vertical axis being the number of pulses:

[0044]

[0045] Where: N is the number of pulses representing the dimension in slow time, and M is the number of cells representing the dimension in fast time.

[0046] S2. Perform a sliding window operation on the data matrix S obtained in step S1 to intercept the original radar echo signal into several sub-data matrices of a set size. In this embodiment, the sliding window operation intercepts a data matrix S of size 64*128 ij = S(i + s:i + s + 128;j + s:j + s + 64), where i and j respectively represent the starting points of the current data in the original cumulative data signal of the radar, s is the sliding size of the sliding window, 64 is the range cell, and 128 is the number of cumulative pulses; then label the target signal and sea clutter signal in the original radar echo signal, label the position of the range cell where the target is located in the data matrix as 1, and label the position of the range cell where the sea clutter is located as 0, and the label data is L ij = L(i + s;j + s:j + s + 64). Cumulate the radar echo data, and finally represent the data matrix S as a data set ({S 11 , S 12 , S 21 ,..., S ij}, i = 1, 2, 3,..., N - s - 128, j = 1, 2, 3,..., M - s - 64) and a label set ({L 11 , L 12 , L 21 ,..., L ij}, i = 1, 2, 3,..., N - s - 128, j = 1, 2, 3,..., M - s - 64). Divide the obtained data set and the corresponding label set into a training set, a validation set, and a test set according to a set ratio;

[0047] S3. Construct a convolutional neural network. Please refer to Figure 2 as shown. The convolutional neural network includes a Focus layer, a first receptive field channel and a second receptive field channel respectively connected to the Focus layer, and a fully connected layer respectively connected to the first receptive field channel and the second receptive field channel, and uses two receptive field channels to extract features in parallel.

[0048] Both the first receptive field channel and the second receptive field channel include:

[0049] A first convolutional layer, a first dimension conversion layer, a multi-head self-attention module, a data stacking and dimension conversion layer, a second convolutional layer, a residual block *2, a second dimension conversion layer, a third convolutional layer, and a third dimension conversion layer connected in sequence.

[0050] The convolutional kernel size of the Focus layer is 3×3, and the stride is 1, which is used to initially extract data features, and then send the data features into the two receptive field channels respectively for data feature extraction.

[0051] The first convolutional layer of the first receptive field channel is a complex convolutional layer with a kernel size of 3×3, a stride of 1×1, and a padding of (1, 1). Then, the data is dimensionally transformed to convert the [batch, 1, m, n] tensor into a tensor of size [m, batch, n]; then it is fed into the complex multi-head self-attention module, and the real part and the imaginary part are output respectively; then the real part and the imaginary part data are stacked on a new dimension and dimensionally transformed to [batch, 2, m, n]. The second convolutional layer is a real convolutional layer with a kernel size of 1×1. The residual block is used to extract data features and prevent overfitting of the data. Then, the data is dimensionally transformed again to convert the [batch, channel, m, n] tensor into [batch, n, m, channel], and then it is fed into the convolutional block with convolutional parameters of a kernel size of 3×1, a stride of 1×1, and a padding of (1, 0). Then, the [batch, 1, m, channel] tensor is dimensionally transformed into [batch, channel, m, 1]. The third convolutional layer is a real convolutional layer with a kernel size of 3×3; finally, the dimension is transformed into [batch, m].

[0052] The first convolutional layer of the second receptive field channel performs a convolutional operation with a kernel size of 3×8, a stride of 1×3, and a padding of (1, 0). Then, the data is dimensionally transformed to convert the [batch, 1, m, n] tensor into a tensor of size [m, batch, n], then it is fed into the complex multi-head self-attention module, and the real part and the imaginary part are output respectively; then the real part and the imaginary part data are stacked on a new dimension and dimensionally transformed to [batch, 2, m, n]. The second convolutional layer is a real convolutional layer with a kernel size of 1×1. The residual block *2 is used to extract data features and prevent overfitting of the data; then the data is dimensionally transformed again to convert the [batch, channel, m, n] tensor into [batch, n, m, channel]; then it is fed into the convolutional block with convolutional parameters of a kernel size of 3×1, a stride of 1×1, and a padding of (1, 0). Then, the [batch, 1, m, channel] tensor is dimensionally transformed into [batch, channel, m, 1]. The third convolutional layer is a real convolutional layer with a kernel size of 3×3; finally, the dimension is transformed into [batch, m].

[0053] In the above description, n and m respectively represent the fast time and slow time of the data. Finally, the output features of the two receptive field channels are combined and fed into the layer layer, and the layer layer parameters are set to 64.

[0054] In two different receptive field channels, by using the second convolutional layer, two residual blocks, a convolutional block, and the third convolutional layer, features are extracted while being compressed. Meanwhile, the multi-head self-attention module can extract diverse features of the data matrix, improve computational efficiency, and enhance the real-time performance of the model. The second convolutional layer and two residual blocks further extract the data features. While the convolutional block extracts the features of the signal on the pulse, it also compresses the features of the signal on the pulse within each convolutional layer. The third convolutional layer realizes feature compression while extracting channel information.

[0055] The design of the two different receptive field channels is to extract features from two different perspectives. The setting of the convolutional parameters in the first receptive field channel enables Channel 1 to capture the local small details in the data matrix, mainly used for extracting the feature information of the sea clutter signal and the target itself, while the second receptive field channel can capture the features in a larger range in the data matrix and can directly capture the overall feature information of the target and the sea clutter.

[0056] S4. Input the training set into the convolutional neural network obtained in step S3, update the weights of the convolutional neural network using the gradient descent method, minimize the loss function, and update the parameters to obtain the trained convolutional neural network. In this embodiment, when using this model for training, the optimizer selects the stochastic gradient descent optimizer Adam.

[0057] In the present invention, the adopted loss function is a cross loss that includes the classification loss of the absence of a target and the regression loss of whether the target is in the correct range cell. This is because if only the classification task is performed on N range cells, overall, the cells occupied by the target in the entire radar echo are not many and are unevenly distributed, making it impossible to locate the position of the target.

[0058] The specific calculation formula of the loss function is:

[0059] Loss total =λLoss ce +(1 - λ)Loss mse

[0060] In the formula: Loss ce is the classification loss for evaluating the absence of a target, Loss mse is the prediction of the target position as the regression loss, and λ is the weight factor for balancing the regression and classification losses.

[0061] When considering the classification problem of the absence of a target and the regression problem of whether the target is in the correct range cell, the cross loss function is used as the classification loss of this model. The calculation formula of the classification loss Loss ce is:

[0062]

[0063] where: N is the number of cells, yi is the label of the i-th range cell, pi is the predicted value of the i-th range cell, and σ(p i ) is the probability that the target exists in the i-th range cell.

[0064] When the prediction of the target position is regarded as a regression problem, when determining that the target is in the correct position, the least mean square error function is used as the classification loss of this model, and the regression loss Loss mse The calculation formula is:

[0065]

[0066] where: y i t is the probability that the target exists at the target position, p i t is the label, and n is the number of targets. That is, the range cells where the target exists are extracted and merged into a vector, and a mean square loss is calculated with the label.

[0067] S5. Input the validation set into the convolutional neural network trained in step S4, set the false alarm rate of the convolutional neural network, and obtain the set threshold through the validation set. In this embodiment, the set false alarm rate is 0.001. After finding the appropriate threshold through the validation set, a radar target signal recognition model is obtained. Input the test set into the radar target signal recognition model, perform target detection on the test set, complete the recognition and positioning of the radar signal, and output the radar target recognition accuracy of the test set.

[0068] The application of the above method will be described in a practical scenario. The radar signal dataset publicly disclosed in Radar Journal is selected. The targets in the dataset are distributed between -10 dB and 15 dB, and there are 10,000 pieces of data in total. The above method is applied to this dataset, and the data is divided into a training set, a validation set, and a test set according to the ratio of 7:2:1. The number of training epochs is set to 200, the batch size is set to 64, the optimizer is set to Adam, the learning rate is set to 0.001. To prevent the model from overfitting, relu_dropout is set to 0.1. To balance the two loss functions, the weight factor is set to 0.5, and an 8-head self-attention mechanism is adopted. A learning rate scheduler ReduceLROnPlateau is created. ReduceLROnPlateau is a scheduler that automatically adjusts the learning rate according to the change of performance metrics during training. When the performance of the model (such as the loss value) does not improve for a period of time, it will reduce the learning rate to help the model continue to optimize. If no improvement in loss is seen for 2 consecutive epochs, the learning rate adjustment will be triggered, and the learning rate will be multiplied by the factor factor, which is set to 0.5 here, so the new learning rate will become 0.0005. Through the above processing, the radar target recognition accuracy of this method on this dataset is obtained. Figure 3 Subfigure (a) of Figure 3 is the amplitude map representation of the model input matrix, Figure 3 and subfigure (b) of Figure 3 is the detection result of the model output.

[0069] On this dataset, we also compared with conventional methods such as CA-CFAR, OS-CFAR, AND, and methods that only use a single channel. The specific results are shown in Table 1 below.

[0070] Table 1

[0071]

[0072] It can be seen from the results that compared with the above comparison methods, the present invention has a significant improvement at any signal-to-noise ratio, ranging from 2% to 25%. The method proposed by the present invention has greater advantages in practical applications and also has good performance in the face of low SCR targets.

[0073] Embodiment 2

[0074] This embodiment provides a radar target signal recognition system based on a self-attention mechanism corresponding to the radar target signal recognition method based on a self-attention mechanism described in Embodiment 1, including the following modules:

[0075] Data matrix construction module: used to collect radar raw echo signals and construct an original data matrix using the radar raw echo signals;

[0076] Data division module: It is used to perform a sliding window operation on the original data matrix, intercept the radar original echo signal into several sub-data matrices of a set size, label the target signal and sea clutter signal in the radar original echo signal, accumulate the radar echo data, and divide the obtained sub-data matrices into a training set, a validation set, and a test set according to a set ratio;

[0077] Model construction module: It is used to construct a convolutional neural network. The convolutional neural network includes a Focus layer, a first receptive field channel and a second receptive field channel respectively connected to the Focus layer, and a fully connected layer respectively connected to the first receptive field channel and the second receptive field channel, and uses two receptive field channels to extract features in parallel;

[0078] Model training module: It is used to input the training set into the convolutional neural network, update the weights of the convolutional neural network by the gradient descent method, minimize the loss function and update the parameters to obtain the trained convolutional neural network; Model validation module: It is used to input the validation set into the trained convolutional neural network, set the false alarm rate of the convolutional neural network, and obtain a set threshold through the validation set to obtain a radar target signal recognition model. Input the test set into the radar target signal recognition model to perform target detection on the test set and output the recognition accuracy of the test set.

Claims

1. A radar target signal recognition method based on self-attention mechanism, characterized in that: The following steps are involved: S1, collecting radar original echo signals, and using the radar original echo signals to construct an original data matrix; S2, performing a sliding window operation on the original data matrix obtained in step S1, intercepting the radar original echo signal into a plurality of sub-data matrices of set sizes, marking the target signal and the sea clutter signal in the radar original echo signal, accumulating the radar echo data, and dividing the obtained sub-data matrices into a training set, a validation set, and a test set according to a set ratio; S3, constructing a convolutional neural network, the convolutional neural network includes a Focus layer, a first receptive field channel and a second receptive field channel respectively connected to the Focus layer, and a fully connected layer respectively connected to the first receptive field channel and the second receptive field channel, and extracting features in parallel using the two receptive field channels; S4, input the training set into the convolutional neural network obtained in step S3, update the weights of the convolutional neural network using the gradient descent method, minimize the loss function and update the parameters to obtain the trained convolutional neural network; S5. Input the verification set into the convolutional neural network trained in step S4, set the false alarm rate of the convolutional neural network, and obtain the set threshold through the verification set to obtain the radar target signal recognition model, input the test set into the radar target signal recognition model, perform target detection on the test set, and output the recognition accuracy of the test set.

2. The radar target signal recognition method according to claim 1, characterized in that: In step S1, the data matrix is ​​a discrete sequence with the horizontal axis being the distance unit cell at fast time and the vertical axis being the number of pulses at slow time.

3. The radar target signal recognition method according to claim 2, characterized in that: In step S2, a sliding window operation is performed on the original data matrix obtained in step S1, and the radar original echo signal is intercepted into a number of 64*128-sized sub-data matrices, where 64 is the distance unit cell under fast time and 128 is the accumulated number of pulses under slow time.

4. The radar target signal recognition method according to claim 1, characterized in that: In step S2, marking the target signal and the sea clutter signal in the original radar echo signal includes: The distance cell position where the target is located in the data matrix is ​​marked as 1, and the distance cell position where the sea clutter is located is marked as 0.

5. The radar target signal recognition method according to claim 1, characterized in that: In step S3, the first receptive field channel and the second receptive field channel both include: The first convolution layer, the first transformation dimension layer, the multi-head self-attention module, the data superposition transformation dimension layer, the second convolution layer, the residual block, the second transformation dimension layer, the third convolution layer, and the third transformation dimension layer are connected in sequence.

6. The radar target signal recognition method according to claim 5, characterized in that: The convolution kernel size of the first convolution layer of the first receptive field channel is 3×3, the step size is 1×1, and the padding is (1, 1); the convolution kernel size of the first convolution layer of the second receptive field channel is 3×8, the step size is 1×3, and the padding is (1, 0).

7. The radar target signal recognition method according to claim 1, characterized in that: In step S4, the loss function is expressed as: Loss total =λLoss ce +(1-λ)Loss mse Where: Loss ce To evaluate the classification loss with or without a target, Loss mse To use the prediction of the target position as regression loss, λ is a weight factor to balance regression and classification losses.

8. The radar target signal recognition method according to claim 7, characterized in that: In the loss function, the classification loss Loss ce The calculation formula is: Where: N is the number of cells, yi is the label of the i-th range cell, pi is the predicted value of the i-th range cell, σ(p i ) is the probability that the target exists in the i-th range unit.

9. The radar target signal recognition method according to claim 7, characterized in that: In the loss function, regression loss Loss mse The calculation formula is: Where: y i t is the probability that a target exists at the target location, p i t is the label and n is the number of targets.

10. A radar target signal recognition system based on self-attention mechanism, characterized in that: include: Data matrix construction module: used to collect radar original echo signals and construct original data matrix using radar original echo signals; Data partitioning module: used to perform sliding window operation on the original data matrix, intercept the original radar echo signal into several sub-data matrices of set size, mark the target signal and sea clutter signal in the original radar echo signal, accumulate radar echo data, and divide the obtained sub-data matrix into training set, verification set and test set according to the set ratio; Model building module: used to build a convolutional neural network, which includes a Focus layer, a first receptive field channel and a second receptive field channel respectively connected to the Focus layer, and a fully connected layer respectively connected to the first receptive field channel and the second receptive field channel, and uses two receptive field channels to extract features in parallel; Model training module: used to input the training set into the convolutional neural network, update the weights of the convolutional neural network using the gradient descent method, minimize the loss function and update the parameters to obtain the trained convolutional neural network; Model verification module: used to input the verification set into the trained convolutional neural network, set the false alarm rate of the convolutional neural network, and obtain the set threshold through the verification set to obtain the radar target signal recognition model, input the test set into the radar target signal recognition model, perform target detection on the test set, and output the recognition accuracy of the test set.

Images