Radar target detection method based on pulse group embedding and sequence analysis

Abstract

The invention provides a radar target detection method based on pulse group embedding and sequence analysis, and the method integrates the induction preference of the radar field into the neural network design through the feature extraction capability of a convolutional network, and replaces the conventional CFAR detection with a self-attention architecture. The pulse group is used as a processing unit, and the method is suitable for radar target detection tasks under various systems.

Description

Technical Field

[0001] The invention belongs to the field of radar information processing, and in particular relates to a radar target detection method based on pulse group embedding and sequence analysis. Background Art

[0002] With the development of artificial intelligence technology, target detection algorithms in the field of computer vision have been fully developed. Target detection architectures such as Yolo and FCOS have shown good results in multiple visual fields. Due to the similarity of data input formats, these detection algorithms in the visual field can be directly applied to radar data. However, there are still some differences between radar data and visual data. It is difficult to fully mine the hidden information in radar data by directly using intelligent algorithms in the visual field. In addition, the current intelligent algorithms in the visual field require the input of the entire image, which does not match the radar working mode in scanning mode. Summary of the invention

[0003] The present invention proposes a radar target detection model based on pulse group embedding and sequence analysis, realizes intelligent target detection applicable to multi-system radars, and can complete high-performance target detection tasks with a low false alarm probability.

[0004] The technical solution to achieve the purpose of the present invention is: a radar target detection method based on pulse group embedding and sequence analysis, comprising:

[0005] Construct a radar pulse group data training set, and use the distance unit position of the target in the pulse group data as the label;

[0006] The radar target detection model is trained using the radar pulse group data training set;

[0007] The radar pulse group data is input into the trained radar target detection model to obtain the range unit position of the target in the pulse group data.

[0008] Preferably, each radar pulse group data includes N pulses.

[0009] Preferably, the radar target detection model comprises a pulse group embedding module and a sequence analysis module.

[0010] Preferably, the pulse group embedding module processes the radar pulse group data in the following manner:

[0011] The pulse group embedding module performs feature mining along the slow time dimension of the pulse group and outputs a pulse group embedding matrix of size B*1*C*M, where B is the number of data samples processed at a single time, 1 represents that the dimension of the slow time dimension in the original input data is reduced to 1 after processing, C represents the number of feature channels, and M is the number of distance units.

[0012] Preferably, the sequence analysis module processes the pulse group embedding matrix output by the pulse group embedding module to form a matrix with a size of B*1*M.

[0013] Preferably, the pulse group embedding module includes a 6-layer one-dimensional complex convolutional neural network. After each convolution layer operation, the second dimension of the data will be half of the original dimension, and the third dimension is the same as the number of convolution layer channels.

[0014] Preferably, the sequence analysis module is a complex Transformer module.

[0015] Preferably, the sequence analysis module outputs a B*1*M matrix, which is a binary matrix, and the two values ​​represent no target and target respectively, and the position of the value of the target represents the distance unit where the target echo is located.

[0016] Compared with the prior art, the present invention has the following significant advantages:

[0017] (1) The present invention uses pulse groups as processing units and is suitable for radar target detection tasks under various systems; the present invention can better suppress sea clutter and retain the ability to detect targets under strong sea clutter background;

[0018] (2) The present invention can process complex data and can completely preserve and transmit the hidden information in the data;

[0019] (3) The information processing flow of the present invention corresponds to the traditional method, which is convenient for subsequent algorithm upgrade and optimization.

[0020] The present invention is described in further detail below in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Flowchart of a radar target detection method based on pulse group embedding and sequence analysis.

[0022] Figure 2 To detect the echo images before and after. DETAILED DESCRIPTION

[0023] The present invention is further described below in conjunction with the accompanying drawings, but the protection scope of the present invention is not limited by the implementation regulations. The present invention proposes a radar target detection model based on pulse group embedding and sequence analysis, which realizes intelligent target detection applicable to multi-system radars and can complete high-performance target detection tasks with low false alarm probability.

[0024] A radar target detection method based on pulse group embedding and sequence analysis, comprising:

[0025] S1: Construct a radar pulse group data training set, and use the distance unit position of the target in the pulse group data as the label;

[0026] The radar pulse group data training set consists of a number of radar pulse group data, and each radar pulse group data consists of N pulses.

[0027] S2: Using the radar pulse group data training set to train the radar target detection model; the radar target detection model includes a pulse group embedding module and a sequence analysis module.

[0028] The radar pulse group data size is B*N*C*M, where B is the data sample size processed at a single time, N is the number of pulse groups, that is, the number of slow time dimension data, C is the number of feature channels, when the input data is a real number, C takes the value of 1, when it is a complex number, C takes the value of 1 or 2, and M is the number of range units, that is, the number of fast time dimension data;

[0029] The technical solution for implementing the present invention is as follows: the input of the model is radar pulse group data consisting of N pulses, and the output is the distance unit position of the target in the pulse group data. The model consists of a pulse group embedding module and a sequence analysis module. The steps of implementing target detection by the model are as follows:

[0030] The radar target detection model processes data as follows:

[0031] The input data is sent to the pulse group embedding module, which performs feature mining along the slow time dimension of the pulse group and outputs a pulse group embedding matrix of size B*1*C*M, where B is the number of data samples processed at a single time, 1 represents that after processing, the dimension of the slow time dimension in the original input data is reduced to 1, C represents the number of feature channels, and M is the number of distance units;

[0032] The sequence analysis module processes each data sample of size 1*1*C*M among the B data samples, that is, the pulse group embedding matrix output by the pulse group embedding module, and the final returned result constitutes a matrix of size B*1*M, where B is the number of data samples processed at a single time and M is the number of distance units.

[0033] In the above steps, taking pure complex radar input data as an example, assuming that the data size is 128*64*1*5000, the above values ​​represent that the amount of data processed in a single time is 128, the number of pulses in a pulse group is 64, the number of initial feature channels is 1, and the number of distance units is 5000.

[0034] In a further embodiment, the pulse group embedding module is composed of a 6-layer one-dimensional complex convolutional neural network. After each convolution layer is operated, the dimension of the second dimension of the data will be half of the original dimension, and the dimension of the third dimension is the same as the number of convolutional layer channels. Assuming that the number of convolutional channels in the last layer of the convolutional neural network is 768, the output pulse group embedding matrix size is 128*1*768*5000.

[0035] In a further embodiment, in the sequence analysis module, for each 1*1*768*5000 embedded sample in the 128 data samples, the sample can be considered as a sequence with a length of 5000, and each sequence element is a 1*768 feature vector. The sequence analysis module is a complex Transformer module, which converts the pulse group embedding matrix with a size of 128*1*768*5000 into a 0-1 matrix with a size of 128*1*5000, wherein the position of 1 in each 1*1*5000 sample represents the distance unit where the target echo is located.

[0036] During the training process, for complex networks, the back propagation algorithm based on Wirtinger calculus is used for training. The objective function of the training adopts the Diceloss formula as shown below:

[0037]

[0038] Among them, X and Y represent the set consisting of model output and label image pixels respectively, and |·| represents the number of set elements.

[0039] S3: Input the radar pulse group data into the trained radar target detection model to obtain the distance unit position of the target in the pulse group data. The sea detection staring mode data is used as the basis for model training. Data labels are made according to the location of the test target given in the data set, and the position where the target exists in the data is marked as 1, and the position where the target does not exist is marked as 2. Due to the limitation of computing resources, on the basis of selecting 16 pulses, only 256 distance units are intercepted to form a 16*1*256 input matrix, which corresponds to the corresponding label. 1 million input matrices are intercepted from the staring mode data together with the corresponding labels to form a training-validation data set. Another set of data is taken as the test set. The model is trained on the NVIDIA RTX-4090 graphics card, 70% of the data samples are selected for training, and 30% of the samples are used for validation. The convergence state is reached after 5 training cycles.

[0040] In the test data, there is a small boat target at the distance unit 1760-1762 under the background of strong sea clutter. A sliding window with a size of 16*1*256 is used to traverse the entire radar image. The detection performance of the converged model is as follows: Figure 2As shown in the figure, it can be seen that under the background of strong sea clutter, the model can suppress sea clutter well and retain the ability to detect targets. According to statistics, the false alarm probability generated by the detection is only 1.8e-3, and the detection probability reaches 99.2%.

[0041] The radar target detection model proposed in the present invention fully integrates the traditional radar information processing flow and the neural network architecture, incorporates the inductive preference in the radar field into the neural network design, and is applicable to radar target detection tasks under various systems.

Claims

1. A radar target detection method based on pulse group embedding and sequence analysis, characterized in that: include: Construct a radar pulse group data training set, and use the distance unit position of the target in the pulse group data as the label; The radar target detection model is trained using the radar pulse group data training set; The radar pulse group data is input into the trained radar target detection model to obtain the range unit position of the target in the pulse group data.

2. The radar target detection method based on pulse group embedding and sequence analysis according to claim 1, characterized in that: Each radar pulse group data includes N pulses.

3. The radar target detection method based on pulse group embedding and sequence analysis according to claim 1, characterized in that: The radar target detection model includes a pulse group embedding module and a sequence analysis module.

4. The radar target detection method based on pulse group embedding and sequence analysis according to claim 3 is characterized in that: The pulse group embedding module processes radar pulse group data as follows: The pulse group embedding module performs feature mining along the slow time dimension of the pulse group and outputs a pulse group embedding matrix of size B*1*C*M, where B is the number of data samples processed at a single time, 1 represents that the dimension of the slow time dimension in the original input data is reduced to 1 after processing, C represents the number of feature channels, and M is the number of distance units.

5. The radar target detection method based on pulse group embedding and sequence analysis according to claim 4 is characterized in that: The sequence analysis module processes the pulse group embedding matrix output by the pulse group embedding module to form a matrix with a size of B*1*M.

6. The radar target detection method based on pulse group embedding and sequence analysis according to claim 3 is characterized in that: The pulse group embedding module includes a 6-layer one-dimensional complex convolutional neural network. After each convolution layer operation, the second dimension of the data will be half of the original dimension, and the third dimension is the same as the number of convolution layer channels.

7. The radar target detection method based on pulse group embedding and sequence analysis according to claim 3 is characterized in that: The sequence analysis module is a complex Transformer module.

8. The radar target detection method based on pulse group embedding and sequence analysis according to claim 3 is characterized in that: The sequence analysis module outputs a B*1*M matrix, which is a binary matrix. The two values ​​represent no target and target respectively, and the position of the value of the target represents the distance unit where the target echo is located.

9. The radar target detection method based on pulse group embedding and sequence analysis according to claim 1, characterized in that: The radar target detection model is trained using the back-propagation algorithm based on Wirtinger calculus.

10. The radar target detection method based on pulse group embedding and sequence analysis according to claim 1, characterized in that: The objective function used to train the radar target detection model is: Among them, X and Y represent the set consisting of model output and label image pixels respectively, and |·| represents the number of set elements.

Images