Radar original domain target identification system based on multi-view antenna distribution design and convolution processing

Abstract

The invention discloses a radar original domain target recognition system based on multi-view antenna distribution design and convolution processing, which is characterized in that a multi-view receiving antenna array is designed to increase receiving information entropy, and a target feature extraction algorithm based on a convolutional neural network is established. According to the method, the strong difference between target characteristics and clutter and multipath interference characteristics is constructed in radar original domain data, so that clutter and multipath interference suppression and other functions in a mainstream radar target recognition normal form are realized; meanwhile, a feature extraction algorithm is designed in combination with the strong feature extraction capability of convolution calculation so as to distinguish the features between the target and clutter, multipath interference and different targets, and finally high-precision radar target recognition is achieved. The method is expected to solve the challenges of complex algorithms such as clutter suppression, coherent accumulation and imaging faced by current high-precision radar target recognition, and provides a new idea for the field of radar target recognition.

Description

Technical Field

[0001] The invention relates to the field of radar target recognition, and in particular to a radar original domain target recognition system based on multi-view antenna distribution design and convolution processing under the condition of clutter and multipath interference. Background Art

[0002] Radar Target Recognition (RTR) is a target detection technology that analyzes radar signals to determine the target type. This process involves extracting key information from radar echoes and matching it with a preset target database to achieve accurate identification and classification of targets.

[0003] Traditional radar target recognition technology relies on the sampling of radar echoes by analog-to-digital converters. After pulse compression and coherent accumulation, target information carriers such as one-dimensional images, two-dimensional images, and range-Doppler images are obtained. Feature extraction and recognition algorithms are then processed based on these information carriers. Current research focuses on the processing of images or other information carriers after radar imaging (A. Deng, N. Qian, S. Hua, J. Wan, Z. Lv, and W. Zou, High-resolution ISAR imaging based on photonic receiving for high-accuracy automatic target recognition, Optics Express, 2022, 30(12): 20580-20588.), while direct processing of the original radar signal is still insufficient.

[0004] In addition, in addition to the target scattered echo, the radar echo is inevitably mixed with clutter and multipath interference, which form false targets in the target information carrier and seriously affect the target recognition results. Therefore, before feature extraction, the radar echo, one-dimensional image or two-dimensional image needs to be preprocessed to remove false targets.

[0005] In order to improve the recognition accuracy, increasing the number of receiving antennas and optimizing their layout have been proven to be effective strategies. Multi-antenna systems can not only capture target echoes from different perspectives and improve azimuth resolution, but also enhance target features through spatial diversity and suppress interference, thereby improving the signal-to-noise ratio. For example, Gerard Lachapelle et al. used array antennas and beamforming technology in 2016 to effectively reduce clutter and multipath interference (reference: Vagle N et al., 2016); the German Aerospace Center used filtering in 2020 to filter out multipath signals using the difference in signal strength received by multiple antennas (reference: Zorn S et al., 2020)

[0006] In addition, multi-angle receiving antennas can further improve detection accuracy. A study by the Italian Naval Research Center in 2018 showed that the more evenly distributed the antennas are and the wider the angle around the target, the higher the azimuth resolution (reference: Lembo L et al., 2018). However, in practical applications, due to the uncertainty of the target position, it is difficult to place the antenna ideally around the target. To this end, the University of Quebec in Canada proposed an adaptive radar receiver placement mechanism in the same year, aiming to maximize the signal-to-noise ratio, reduce the bit error rate, and improve positioning accuracy (reference: Kilani MB et al., 2018).

[0007] Although there have been many studies on optimizing the received signal by changing the distribution of receiving antennas, the following challenges still exist: although multi-antenna reception can improve the signal-to-noise ratio, subsequent digital signal processing is still required to completely suppress noise and interference; existing research has mostly focused on simple radar functions, such as early warning and positioning, while research on complex radar target recognition functions is still insufficient; at the same time, methods such as beamforming still require post-processing of radar raw data when dealing with clutter and multipath interference, which increases processing delay and power consumption. Summary of the invention

[0008] In view of the above-mentioned deficiencies in the prior art, the present invention provides a radar original domain target recognition system based on multi-view antenna distribution design and convolution processing in the case of clutter and multipath interference, increases the received information entropy by designing a multi-view receiving antenna array, and establishes a target feature extraction algorithm based on convolution calculation. After the signal echo containing clutter and multipath interference is received by the multi-view antenna array, the signals at different angles interact with each other in the feature extraction module based on convolution calculation to highlight the signal features, so as to achieve the purpose of removing clutter and multipath interference, and finally achieve high-precision radar target recognition.

[0009] The technical solution of the present invention is as follows:

[0010] A radar original domain target recognition system based on multi-view antenna distribution design and convolution processing, which has the following characteristics:

[0011] A waveform transmitter module for generating and transmitting radar signals;

[0012] A multi-view antenna array receiver module, including multiple receiving antenna arrays and multiple analog-to-digital converters, for receiving radar signals from multiple viewpoints and converting them into digital signals;

[0013] a single antenna receiver module, including a receiving antenna and an analog-to-digital converter for receiving radar signals from space and converting them into digital signals;

[0014] A convolutional neural network-based feature extractor module for extracting high-dimensional, abstract features from data output by the multi-view antenna array receiver module and the single antenna receiver module;

[0015] A recognition classifier module is used to classify and recognize the target based on the extracted features;

[0016] The decision module is used to compare the recognition results of the network trained under different antenna configurations and trigger the feedback optimization module;

[0017] A feedback optimizer module, used to analyze the receiving performance of the multi-view antenna array receiver according to the output of the decision module, and optimize the arrangement array of the multi-view receiving antenna;

[0018] Among them, the output end of the multi-view antenna array receiver module and the output end of the single antenna receiver module are respectively connected to the input end of the feature extractor module based on the convolutional neural network, the output end of the feature extractor module based on the convolutional neural network is connected to the input end of the recognition classifier module, the output end of the recognition classifier module is connected to the input end of the decision module, the output end of the decision module is connected to the input end of the feedback optimizer module, and the output end of the feedback optimizer module is connected to the input end of the multi-view antenna array receiver module to form a closed-loop optimization system.

[0019] The waveform transmitter is a broadband signal transmitting device based on multiple or single antennas.

[0020] The multi-viewing angle antenna array receiver is a radar data receiving module based on multiple receiving antenna arrays.

[0021] The analog-to-digital converter in the multi-viewing angle antenna array receiver is a data acquisition board or an oscilloscope.

[0022] The single-antenna receiver is a radar data receiving module based on a single receiving antenna.

[0023] The feature extractor based on convolutional neural network is a computer program that implements the convolutional network.

[0024] The recognition classifier is a computer program based on nonlinear calculation and fully connected layers.

[0025] The feedback optimizer is a computer program that implements algorithms for various multi-antenna array configurations based on various receiving performances such as anti-interference performance and target feature extraction capability.

[0026] The radar original domain target recognition technology based on multi-view antenna distribution design and convolution processing is used to realize a high-precision radar target recognition method. The method includes two stages: training and applying a feature extractor based on a convolutional neural network and a recognition classifier. The steps are as follows:

[0027] 1) Training phase:

[0028] The feature extractor and recognition classifier based on the convolutional neural network can only realize high-precision recognition of the radar original domain signals output by the multi-view antenna array receiver and the single antenna receiver after training. After the waveform transmitter transmits the broadband signal, the signal propagates through space, and the radar echo with clutter and multipath interference is received by the multi-view antenna array and the single antenna receiver respectively, and two sets of radar original domain data with different antenna configurations are obtained. The obtained radar original domain data is used as the training set of the feature extractor and recognition classifier based on the convolutional neural network, and the feature extractor and recognition classifier based on the convolutional neural network are trained using an optimization algorithm, so that the feature extractor and recognition classifier based on the convolutional neural network can learn the features between different targets and between targets and clutter and multipath interference, and realize classification and recognition. At the same time, the decision device compares the recognition results of the network trained by the data sets obtained according to different antenna configurations. If the recognition accuracy of the network trained by the data received by the multi-view antenna array receiver is lower than the recognition accuracy of the network trained by the data received by the single antenna receiver, the feedback optimizer will analyze the receiving performance of the multi-view antenna array receiver and re-optimize the arrangement array of the multi-view receiving antenna according to the algorithm. Otherwise, the training is completed, and finally high-precision radar target recognition is achieved.

[0029] 2) Application phase:

[0030] After the waveform transmitter transmits the broadband signal, the multi-view antenna array receiver receives the radar echo with clutter and multipath interference, etc., which has been propagated through space. The received signal is then input into the feature extractor and recognition classifier based on the convolutional neural network trained in the training phase, and the network outputs a high-accuracy recognition accuracy to achieve high-precision radar target recognition.

[0031] Based on the above technical features, the present invention has the following advantages:

[0032] 1. Use a multi-view antenna array to receive radar signals contaminated by clutter and multipath interference. Through multi-view reception combined with a feature extractor based on a convolutional neural network, the target features are amplified to suppress clutter and multipath interference.

[0033] 2. The feature extractor and recognition classifier based on convolutional neural network is an end-to-end, simple, fast and efficient radar target recognition method based on target feature learning. Its characteristic is that it directly processes the radar original domain signal and avoids complicated signal processing procedures.

[0034] The present invention constructs a strong difference between target characteristics and clutter and multipath interference characteristics based on a multi-view antenna array design, forming an innovative method for suppressing clutter and multipath interference, thereby receiving contaminated radar echoes in the electromagnetic original domain, and then directly performing feature extraction based on a convolutional neural network to obtain the key features of the target, and then executing a classification and recognition algorithm to achieve high-precision target recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0036] Figure 1 This is a module composition diagram of a radar original domain target recognition system based on multi-view antenna distribution design and convolution processing in one embodiment of the present invention.

[0037] Figure 2 It is a processing flow chart of the feedback optimization module in one embodiment of the present invention.

[0038] Figure 3 It is a structural schematic diagram of a feature extractor module and a recognition classifier module based on a convolutional neural network in one embodiment of the present invention.

[0039] Figure 4 This figure shows the training effect of a feature extractor module and a recognition classifier module based on a convolutional neural network in one embodiment of the present invention. DETAILED DESCRIPTION

[0040] The following is a detailed description of an embodiment of the radar original domain target recognition technology based on multi-view antenna distribution design and convolution processing of the present invention: This embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation method and specific operation process are given, but the protection scope of the present invention should not be limited to the following embodiments.

[0041] Figure 1 This is a module composition diagram of a radar original domain target recognition system based on multi-view antenna distribution design and convolution processing in one embodiment of the present invention. The present invention mainly includes the following operations:

[0042] Design the placement of each antenna in a multi-antenna receiving array;

[0043] The transmitting antenna transmits a signal to the target, and each receiving antenna receives the signal echo contaminated by clutter and multipath interference, and then processes the data to obtain the joint received data;

[0044] Set up a single receiving antenna to receive the echo of the target detected by the transmitting antenna under the same conditions;

[0045] A convolution-based neural network is used to process the joint reception data obtained by the multi-antenna receiving array and the data received by a single antenna respectively;

[0046] Compare and judge the recognition results of the data obtained by the multi-antenna receiving array and the single receiving antenna;

[0047] If the recognition effect of the multi-antenna receiving array is worse or does not meet the accuracy requirements, the spatial arrangement of the multi-antenna receiving array is re-optimized, and conversely the array is used to achieve high-precision recognition of the target.

[0048] like Figure 1 As shown, the radar original domain target recognition system based on multi-view antenna distribution design and convolution processing has a framework structure including: waveform transmitter module 1, multi-view antenna array receiver module 2, single antenna receiver module 3, feature extractor module 4 based on convolutional neural network, recognition classifier module 5, decision module 6 and feedback optimizer module 7. Its construction process is divided into two stages: training stage and application stage. In the training stage, we optimize the multi-view antenna array receiver module 2 and train the feature extractor module 4 based on convolutional neural network and the recognition classifier module 5, so that it can recognize the target with high precision. In the application stage, we use the multi-view antenna array receiver module 2 optimized in the training stage and the feature extractor module 4 based on convolutional neural network and the recognition classifier module 5 trained to recognize radar targets with high precision. In the embodiment, the waveform transmitter module 1 uses an arbitrary waveform generator to generate a linear frequency modulation signal with a bandwidth of 1GHz, which is then modulated to a frequency band of 10GHz by a mixer, and the signal is transmitted into space through an antenna. The receiving antenna of the multi-view antenna array receiver module 2 is placed at 90° with the line connecting the transmitter and the target as the symmetry axis. The receiving antenna in the single antenna receiver module 3 is placed next to the transmitting antenna. The feature extractor module 4 based on convolutional neural network is a convolutional neural network, and the recognition classifier module 5 is a fully connected layer. The structural diagram is shown in FIG. Figure 3As shown. As can be seen from the figure, the feature extractor module 4 based on the convolutional neural network includes an input layer, two one-dimensional convolution layers and two maximum pooling layers; the recognition classifier module 5 is composed of a fully connected layer and an output layer. The above two modules used in the embodiment are implemented in a personal computer. The judge module 6 is a comparator based on recognition accuracy implemented in a personal computer. It judges whether the receiving antenna arrangement in the multi-view antenna array receiver module 2 is optimal by comparing the data obtained by the multi-view antenna array receiver module 2 and the data obtained by the single antenna receiver module 3 after training. The processing flow chart of the feedback optimizer module 7 is shown as follows. Figure 2 As shown, the simulated annealing algorithm is used to determine the optimal position of the antenna arrangement. The objective function (system energy) is jointly determined by the peak sidelobe ratio PSLR of the received signal and the half-peak full width FWHM of the target peak lobe, where PSLR corresponds to the intensity of the signal received by the system being interfered by clutter and multipath. The higher the PSLR, the lower the clutter and multipath interference to the signal; FWHM corresponds to the clarity of the radar system detecting the target. The lower the FWHM, the narrower the target peak lobe, the more scattering points can be detected, and the corresponding azimuth resolution is larger. First, set the initial coefficients a = 1, b = -1, and set the objective function (system energy) E = a·PSLR+b·FWHM. The simulated annealing algorithm can be used to calculate the optimal antenna arrangement position at this time. If it is considered that the current antenna arrangement causes the recognition accuracy to be too low after the decision module 6, it is analyzed whether the low recognition accuracy is caused by clutter, multipath interference (case 1) or unclear target features (case 2). For case 1, increase the weight of the peak sidelobe ratio PSLR in the objective function, that is, increase a; for case 2, increase the weight of the full width at half maximum FWHM in the objective function, that is, reduce b (increase the absolute value of b)

[0049] The module connections and module functions of the two stages in the construction process are described as follows.

[0050] 1) Training phase:

[0051] First, the microwave source generates a 10GHz carrier signal, and the arbitrary waveform generator generates a linear frequency modulation signal with a bandwidth of 1GHz, which is modulated to a carrier frequency of 10GHz by a mixer. Second, the number of receiving antennas is set to 2, and the receiving antenna of the multi-view antenna array receiver module 2 is placed at 90° with the line connecting the transmitter and the target as the symmetry axis. Third, two different types of identification targets, circular and triangular, are set, and the antenna can detect target features at different angles and obtain data sets by rotating the target. For each type of target, a target echo is obtained every 0.18°, then a total of 4000 sets of data are obtained for the two types of targets, and the training set and the test set are randomly divided according to the ratio of 3:1. Fourth, the collected training set is input into the feature extractor module 4 based on the convolutional neural network and the recognition classifier module 5 for training. Among them, for the data obtained by the single antenna receiver module 3, the received data of the two antennas are spliced ​​to obtain a one-dimensional vector. For the single antenna receiver module 3, since there is only one antenna receiving data, in order to ensure that the input data dimension of the neural network is the same, the single antenna receiving data is copied and spliced ​​to obtain a one-dimensional vector. The Adam optimization algorithm is selected to reduce the absolute error between the network input and the network reference, and the learning rate of the optimization algorithm is 0.0001. During the 30,000 iterations, the network parameters are continuously adjusted, and finally the feature extractor module 4 based on the convolutional neural network and the recognition classifier module 5 after training can learn the characteristics of different targets and achieve high-precision radar target recognition.

[0052] Figure 4 The figure is a training effect diagram of the feature extractor module 4 based on the convolutional neural network and the recognition classifier module 5. Among them, Figure (a) is a training effect diagram obtained by collecting data from the multi-view antenna array receiver module 2 when the number of receiving antennas is 2, and Figure (b) is a training effect diagram obtained by collecting data from the single antenna receiver module 3. It can be seen that when the network is trained with the data of the multi-view antenna array receiver module 2, the network loss can be reduced to 0.0344 at the lowest, and the network converges quickly; while when the network is trained with the data of the single antenna receiver module 3, the network loss can only be reduced to about 0.159 at the lowest, and the network converges slowly, and is close to convergence at the end of training.

[0053] 2) Application phase:

[0054] After the waveform transmitter transmits the broadband signal, the multi-view antenna array receiver receives the radar echo with clutter and multipath interference, etc., which has propagated through space. The received signal is then input into the feature extractor and recognition classifier based on the convolutional neural network trained in the training phase. For 1,000 test sets, the recognition accuracy of the network output layer is about 99.4%, achieving high-precision radar target recognition.

[0055] All matters not covered in the above embodiments of the present invention are well known in the art.

[0056] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art may make various modifications or variations within the scope of the claims, which do not affect the essence of the present invention.

Claims

1. A radar original domain target recognition system based on multi-view antenna distribution design and convolution processing, characterized in that: include: A waveform transmitter module (1) for generating and transmitting a radar signal; A multi-viewing angle antenna array receiver module (2), comprising a plurality of receiving antenna arrays and a plurality of analog-to-digital converters, for receiving radar signals from a plurality of viewing angles and converting them into digital signals; A single antenna receiver module (3), comprising a receiving antenna and an analog-to-digital converter, for receiving radar signals from space and converting them into digital signals; A feature extractor module (4) based on a convolutional neural network, used to extract high-dimensional, abstract features from the data output by the multi-view antenna array receiver module (2) and the single antenna receiver module (3); A recognition classifier module (5), used for classifying and identifying the target based on the extracted features; A decision module (6) is used to compare the recognition results of the networks trained under different antenna configurations and trigger the feedback optimization module (7); A feedback optimizer module (7) is used to analyze the receiving performance of the multi-view antenna array receiver according to the output of the decision module (6), and optimize the arrangement array of the multi-view receiving antenna; The output end of the multi-view antenna array receiver module (2) and the output end of the single antenna receiver module (3) are respectively connected to the input end of the feature extractor module (4) based on the convolutional neural network, the output end of the feature extractor module (4) based on the convolutional neural network is connected to the input end of the recognition classifier module (5), the output end of the recognition classifier module (5) is connected to the input end of the decision module (6), the output end of the decision module (6) is connected to the input end of the feedback optimizer module (7), and the output end of the feedback optimizer module (7) is connected to the input end of the multi-view antenna array receiver module (2), so as to form a closed-loop optimization system.

2. The radar original domain target recognition system based on multi-view antenna distribution design and convolution processing according to claim 1 is characterized in that: The waveform transmitter module (1) is a broadband signal transmitting device based on multiple or single antennas.

3. The radar original domain target recognition system based on multi-view antenna distribution design and convolution processing according to claim 1 or 2 is characterized in that: The waveform transmitter module (1) comprises an arbitrary waveform generator for generating a radar baseband signal, a microwave source for generating a high-frequency carrier signal, a mixer for up-converting the radar baseband signal to a carrier frequency, an amplifier for increasing the signal power in a microwave link, and a transmitting antenna for transmitting the radar signal into space.

4. The radar original domain target recognition system based on multi-view antenna distribution design and convolution processing according to claim 1 is characterized in that: The multi-viewing angle antenna array receiver module (2) is a radar data receiving module based on multiple receiving antenna arrays.

5. The radar original domain target recognition system based on multi-view antenna distribution design and convolution processing according to claim 4 is characterized in that: The analog-to-digital converter in the multi-viewing angle antenna array receiver module (2) is a data acquisition board or an oscilloscope.

6. The radar original domain target recognition system based on multi-view antenna distribution design and convolution processing according to claim 1 is characterized in that: The single antenna receiver module (3) is a radar data receiving module based on a single receiving antenna.

7. The radar original domain target recognition system based on multi-view antenna distribution design and convolution processing according to claim 1 is characterized in that: The feature extractor module (4) based on convolutional neural network is a computer program for realizing convolutional neural network; the recognition classifier module (5) is a computer program based on nonlinear calculation and full connection layer; the feedback optimizer module (7) is a computer program for realizing algorithms for setting various multi-antenna arrays based on various receiving performances such as anti-interference performance and target feature extraction capability.

8. A method for realizing high-precision radar target recognition by using the radar original domain target recognition system based on multi-view antenna distribution design and convolution processing as described in any one of claims 1 to 7, characterized in that: The steps include: The waveform transmitter transmits broadband signals, and the multi-view antenna array receiver receives radar echoes and inputs the trained feature extractor and recognition classifier based on convolutional neural network to achieve high-precision radar target recognition. The training of the feature extractor and recognition classifier based on the convolutional neural network includes: After the waveform transmitter module (1) transmits the broadband signal, the signal propagates through space, and the radar echo with clutter and multipath interference is received by the multi-view antenna array receiving module (2) and the single antenna receiver module (3) respectively, so as to obtain two groups of radar original domain data with different antenna configurations; The radar original domain data is used as a training set for the feature extractor module (4) based on the convolutional neural network and the recognition classifier module (5), and the feature extractor module (4) based on the convolutional neural network and the recognition classifier module (5) are trained using an optimization algorithm, so that the feature extractor module (4) based on the convolutional neural network and the recognition classifier module (5) can learn the features between different targets and between the targets and clutter and multipath interference, thereby achieving classification and recognition; The decision module (6) compares the recognition results of the network trained with the data sets obtained according to different antenna configurations. If the recognition accuracy of the network trained with the data received by the multi-view antenna array receiver module (2) is lower than the recognition accuracy of the network trained with the data received by the single antenna receiver module (3), the feedback optimizer module (7) will analyze the receiving performance of the multi-view antenna array receiver module (2) and re-optimize the arrangement array of the multi-view receiving antenna according to the algorithm; otherwise, the training is completed, and finally high-precision radar target recognition is achieved.

Images