JP7577613B2 - Radar system and radar signal processing method - Google Patents

Description

This embodiment relates to a radar system and a radar signal processing method.

In conventional radar systems, when detecting small targets in environments with a lot of clutter and thermal noise, the signal-to-noise ratio (SNR) is low, leading to problems such as failure to detect targets or false detections, and large angle measurement errors.

Pulse Compression, Yoshida, 'Revised Radar Technology', Institute of Electronics, Information and Communication Engineers, pp. 278-280 (1996) CFAR (Constant False Alarm Rate), Yoshida, 'Revised Radar Technology', Institute of Electronics, Information and Communication Engineers, pp.87-89 (1996) Monopulse, Yoshida, 'Revised Radar Technology', Institute of Electronics, Information and Communication Engineers, pp. 260-264 (1996) Convolutional Neural Network (CNN), Saito, 'Deep Learning from Scratch', O'Reilly Japan, pp.205-221 (2016) Denoising, Kai Zhang, ‘Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising’, IEEE Transactions on Image Processing, Vol.26, Issue 7, July (2017) SRCNN, Chao Dong, ‘Image Super-Resolution Using Deep Convolutional Networks’, arXiv:1501.00092v3, July (2015)

As described above, when detecting small targets in environments with a lot of clutter and thermal noise, conventional radar systems have a low signal-to-noise ratio (SNR), which can result in failure to detect targets or false detections, and large angle measurement errors.

The objective of this embodiment is to provide a radar system and radar signal processing method that can reduce false positives, detect small targets, and improve angle measurement accuracy, even in clutter and thermal noise environments.

In order to solve the above problems, the radar system according to this embodiment is a radar system that performs coherent integration processing on the slow-time axis using a single pulse or modulated N (N≧1) pulses, and is equipped with a learning means and a processing means. The learning means acquires range-Doppler data (RD data) in advance, and while overlapping the RD data on the range axis and the Doppler axis, trains a CNN (Convolutional Neural Network) model using learning data that combines unwanted wave data divided into M ways and previously acquired target data (teacher signal). The processing means uses the learning model to divide the RD data of the Σ beam and the squint beam of the AZ (Azimuth) axis (EL (Elevation) axis) input in real time into M ways, and uses the RD data synthesized by denoising for each division unit to detect targets using the Σ beam through target detection processing based on CFAR (Constant False Alarm Rate) or maximum values, etc., and performs angle measurement processing using the Σ beam and the squint beam of the AZ axis (EL axis).

FIG. 1 is a block diagram showing the configuration of a transmission system of a radar system according to a first embodiment. FIG. 2 is a block diagram showing the configuration of a receiving system of the radar system according to the first embodiment. FIG. 3 is a conceptual diagram showing the denoising process of the reception system shown in FIG. FIG. 4 is a flowchart showing a procedure for executing the denoising process shown in FIG. 3 by inference processing using a learning model. FIG. 5 is a diagram showing learning data and teacher data used in the learning model of FIG. FIG. 6 is a flowchart showing the flow of the learning process of the learning model used in the inference process of FIG. FIG. 7 is a diagram for explaining the denoising process shown in FIG. 3 by comparing before and after the learning model process. FIG. 8 is a diagram showing a method of calculating an error voltage ε using a Σ beam and a Δ beam and measuring an angle using a previously acquired error voltage table in the first embodiment. FIG. 9 is a diagram showing a method of calculating an error voltage ε by squint angle measurement of the Σ beam and the Σ2 beam, and measuring the angle using a previously acquired error voltage table, in the first embodiment. FIG. 10 is a diagram showing an example of performing denoising processing on each of the Σ beam, the Σ AZ beam, and the Σ EL beam in the first embodiment. FIG. 11 is a block diagram showing the configuration of a receiving system of a radar system according to the second embodiment. FIG. 12 is a flowchart showing the flow of the denoising process according to the second embodiment. FIG. 13 is a conceptual diagram specifically illustrating the denoising process according to the second embodiment. FIG. 14 is a block diagram showing the configuration of a receiving system of a radar system according to the third embodiment. FIG. 15 is a flowchart showing a flow of the denoising process according to the third embodiment when two types of learning models are included. FIG. 16 is a flowchart showing a flow of the denoising process according to the third embodiment when one type of learning model is included. FIG. 17 is a block diagram showing the configuration of a receiving system of a radar system according to the fourth embodiment. FIG. 18 is a diagram showing an example in which the antenna aperture is divided into subarrays in the fourth embodiment. FIG. 19 is a block diagram showing an example of the configuration of a sub-array type DBF applied to the fourth embodiment. FIG. 20 is a diagram showing a coordinate system for explaining phases for beam scanning of reception signals of subarrays in the fourth embodiment.

The following describes the embodiment with reference to the drawings.

First Embodiment Denoising Processing FIG. 1 is a block diagram showing the configuration of a transmission system of a radar system according to a first embodiment, and FIG. 2 is a block diagram showing the configuration of the reception system thereof.

In the transmission system shown in FIG. 1, a signal generator 11 generates a transmission seed signal, a modulator 12 modulates and multiplexes transmission information onto the transmission seed signal, a frequency converter 13 converts the modulated signal into a high-frequency signal, a pulse modulator 14 pulse-modulates the high-frequency signal to generate a transmission pulse train, and a transmission antenna 15 transmits N (N≧2) hit pulses.

Next, the receiving system shown in Figure 2 will be described. In the receiving system, the receiving antenna 21 receives the signals of the Σ beam, the Σ AZ beam squinted on the AZ axis, and the Σ EL beam squinted on the EL axis, and each received signal is frequency converted by the frequency converter 22 and converted to a digital signal by the AD converter 23. Next, the FFT/PC (Fast Fourier Transform/Pulse Compression) processor 24 performs FFT on the fast-time axis, and if the radar signal is a pulse compressed signal (range compression, see Non-Patent Document 1), it multiplies it by a reference signal on the range-frequency axis to obtain range-Doppler data (RD data).

Next, the DBF (Digital Beam Forming) processor 25 forms signals for the Σ beam, Σ AZ beam, and Σ EL beam, and the denoising detector 26 sequentially extracts RD data from the received signals for each of the Σ beam, Σ AZ beam, and Σ EL beam at a predetermined timing, decomposes the data into predetermined sizes on the range axis and Doppler axis, synthesizes the data in units of division, extracts maximum values that exceed a predetermined threshold, detects targets, and outputs the target values as observation values from the observation value output device 27. Here, CFAR (Constant False Alarm Rate, see Non-Patent Document 4) may be used instead of extreme value detection.

Meanwhile, the target detection results from the denoising detector 26 are sent to the cell extractor 28, which compares the received signals of the Σ beam, ΣAZ beam, and ΣEL beam formed by the DBF processor 25 to extract cells that have detected targets, and the denoising angle measurer 29 performs angle measurement processing using the Σ beam, ΣAZ beam, and ΣEL beam. The angle measurement processing results are sent to the observation value output device 27, which outputs the observation value of the target detection.

The processing operations of the radar system configured as above will be explained with reference to Figures 3 to 10.

Here, FIG. 3 is a conceptual diagram showing the denoising process of the receiving system shown in FIG. 2, and FIG. 4 is a flowchart showing the procedure for performing the denoising process shown in FIG. 3 by inference processing using a learning model.

First, in the transmission system, as shown in FIG. 1, a transmission seed signal is generated by a signal generator 11, modulated by a modulator 12, converted into a high-frequency signal by a frequency converter 13, pulse-modulated by a pulse modulator 14, and transmitted from a transmission antenna 15 as a pulse with N (N≧2) hits.

Meanwhile, in the receiving system, as shown in FIG. 2, the signal received by the receiving antenna 21 is frequency converted by the frequency converter 22 and converted to a digital signal by the AD converter 23. Next, the signal is converted to the frequency domain by FFT processing on the slow-time axis by the FFT/PC processor 24, and in the case of a pulse-compressed (Non-Patent Document 1) signal, it is pulse-compressed (range-compressed, Non-Patent Document 1) to obtain range-Doppler data (RD data).

This RD data is extracted sequentially at a specified timing by the denoising detector 26, and is decomposed into a specified size on the range axis and Doppler axis. This is shown in Figure 3. As shown in Figure 3, when the detection area of the RD data is subdivided, if a target is located near the boundary of the subdivided area, it may not be possible to detect it. For this reason, the subdivided units are made to overlap on the range-Doppler axis.

In the denoising detector 26, the process shown in FIG. 4 is performed for each of the RD data of the small division units, the division units are data-combined, and the maximum value exceeding a predetermined threshold is extracted to detect the target. The denoising detection process shown in FIG. 4 first divides the data into the small division units (step S11), executes an inference process using a learning model (step S12), determines whether the division unit process is completed (step S13), and if not, changes to the process of the next division unit (step S14) and returns to the process of step S11. Also, if the division unit process is completed in step S13, the data obtained in the division unit is combined (step S15), and the target is detected from the combined data (step S16), and the series of processes is terminated. In step S16, extreme value detection may be used to detect the target, or CFAR (see Non-Patent Document 2) may be used instead.

Now, the learning method of the learning model used in the inference process of step S12 will be described with reference to Figs. 5 to 7. Figs. 5(a) and (b) respectively show the learning data and teacher data used in the learning model of Fig. 4. Fig. 6 shows the learning process flow of the learning model used in the inference process of Fig. 4. Figs. 7(a) and (b) show the denoising process shown in Fig. 3, comparing before and after the learning model process.

As shown in FIG. 5(a), the learning data is created by superimposing a target signal on data obtained by dividing the RD data into small parts. As the target signal, actual measurements or calculated values are prepared in advance, and the cell position on the range-Doppler axis, amplitude, width of the target signal, etc. are set randomly and superimposed on each small division unit. This makes it possible to generate learning data S+N in which the target signal S is superimposed on unnecessary signals N such as clutter and thermal noise of actual measurements. As shown in FIG. 5(b), the teacher data is generated by setting only the target S used in the learning data for each small division unit. Using this learning data, teacher data, and a learning model, learning is performed using a learning model consisting of a CNN (convolutional neural network, non-patent document 4) or the like.

Specifically, as shown in FIG. 6, learning data S+N is generated by superimposing a target signal S on unnecessary signals N of actual measured values of clutter and thermal noise (step S21), a learning model is set (step S22), learning processing is performed using the learning data S+N and teacher data S consisting only of the target signal S (step S23), and it is determined whether a predetermined number of epochs has been completed (step S24). If not, the learning data is changed and the process returns to step S21. If the number of epochs has been completed, the learning parameters of the learning model at the time of completion are saved and the series of processes ends. The above process is called denoising processing (Non-Patent Document 5).

The above learning process is usually performed offline, but in real-time processing, inference processing is performed using the model of the learning results. As a result of learning, it is possible to suppress false positives and extract the target, as is clear from comparing the situation before the learning model processing shown in Figure 7(a) with the situation after the learning model processing shown in Figure 7(b).

The above is the processing of the Σ beam (sum beam) by the receiving antenna 21. Next, we will explain this with reference to Figures 8 to 10.

Here, Figures 8(a) and (b) show a method of calculating error voltage ε using the Σ beam and Δ beam and measuring the angle using a previously acquired error voltage table, Figures 9(a) and (b) show a method of calculating error voltage ε using squint angle measurement of the Σ beam and Σ2 beam and measuring the angle using a previously acquired error voltage table, and Figures 10(a), (b), and (c) show examples of denoising processing for each of the Σ beam, ΣAZ beam, and ΣEL beam.

First, to measure the angle, phase monopulse angle measurement is usually used, using a Δ beam (difference beam) formed by dividing the antenna aperture into two on the AZ axis (EL axis) (see Non-Patent Document 3). As shown in Figures 8(a) and (b), this is a method of measuring the angle by calculating the error voltage ε shown in the following formula using the Σ beam and Δ beam, and using a previously acquired error voltage table.

When denoising is performed, the phase component is lost, so this phase monopulse processing cannot be applied. As a countermeasure, a squint angle measurement is applied that uses only the amplitude component of a squint beam ΣAZ (ΣEL) that has an angle shifted within the beam width on the AZ axis (EL axis) for the Σ beam. This is shown in Figure 9.

To perform this processing, denoising is performed on each of the Σ beam, ΣAZ beam, and ΣEL beam, as shown in Figures 10(a), (b), and (c). Target detection processing is performed using the output of the Σ beam after denoising processing, target detection cells are extracted from the Σ beam, ΣAZ beam, and ΣEL beam, and angle measurement processing is performed using equation (2). With this method, angles can be measured using signals in which unnecessary signals have been suppressed by denoising processing, improving angle accuracy.

As described above, the radar system according to this embodiment is a radar system that performs coherent integration processing on the slow-time axis using a single pulse or modulated N (N≧1) pulses, acquires range-Doppler data (RD data) in advance, and while overlapping the RD data on the range axis and Doppler axis, uses learning data that combines unwanted wave data divided into M ways and previously acquired target data (teacher signal) to train a CNN (Convolutional Neural Network) model. Using this learning model, RD data of the Σ beam and squint beam of the AZ axis (EL axis) that is input in real time is divided into M ways, and using the RD data that is denoised and combined for each division unit, the Σ beam is used to detect targets by target detection processing using CFAR (Constant False Alarm Rate) or maximum values, etc., and angle measurement processing is performed using the Σ beam and the squint beam of the AZ axis (EL axis).

In other words, according to this embodiment, the effects of clutter and thermal noise can be suppressed by denoising processing, which makes it possible to detect small targets and improve angle measurement accuracy.

Second embodiment: Pre-processing In the first embodiment, a method for measuring an angle by suppressing unnecessary waves using a learning model process as denoising processing has been described. In this embodiment, a method for improving the angle measurement accuracy will be described with reference to Figs. 11 to 13. In Figs. 11 and 12, the same parts as those in Figs. 2 and 3 are denoted by the same reference numerals, and only the different parts will be described here.

Figure 11 is a block diagram showing the configuration of a receiving system according to the second embodiment, Figure 12 is a flowchart showing the flow of the denoising process shown in Figure 11, and Figure 13 is a conceptual diagram specifically showing the denoising process shown in Figure 12.

The second embodiment differs from the first embodiment in that a pre-processor 2A is placed before the denoising detection processor 26, and pre-processing (step S17) is performed before the inference process (step S12) in the denoising process.

The above pre-processing is shown in Figure 13. First, the rms (root mean square) value of the noise is calculated using data from the RD data at long distances and near high Doppler, where there is a low possibility of a target being present, and the minimum signal value is calculated by multiplying it by a specified coefficient. Next, the range-Doppler data of the Σ beam and the ΣAZ (ΣEL) beam is divided. Next, for each division unit of the Σ beam, if the maximum amplitude exceeds the minimum signal value, it is normalized by the maximum amplitude value. The minimum signal value is set to prevent the signal amplitude from becoming large due to normalization when there is only noise. In addition, normalization is performed to prevent variation in the unwanted wave suppression effect of the denoise due to the magnitude of the signal amplitude.

This normalization coefficient is used to normalize ΣAZ (ΣEL). As a result, even after normalization, the amplitude ratio between the Σ beam and the ΣAZ (ΣEL) beam is preserved, making it possible to perform squint angle measurement using equations (3) and (4).

As described above, according to this embodiment, a normalization coefficient is calculated for each divided data by the maximum amplitude of the Σ beam, and the same normalization coefficient is applied to the Σ beam and the squint beam to perform denoising. In other words, normalizing each divided data by the maximum value of the Σ beam suppresses variation in the suppression effect due to the magnitude of the signal amplitude, and applying the normalization coefficient of the Σ beam to the squint beam preserves the difference between the Σ beam and the squint beam, thereby enabling highly accurate squint angle measurement.

(Third embodiment) Shortcut In the first embodiment, a general configuration using a CNN or the like as a learning model was described. In this embodiment, a denoising method (Non-patent 5) using a shortcut with a high effect of suppressing unnecessary waves will be described with reference to Figs. 14 to 16. Fig. 14 is a block diagram showing the configuration of a receiving system of a radar system according to a third embodiment, Fig. 15 is a flowchart showing a flow of a denoising process in the third embodiment in which two types of learning models are included, and Fig. 16 is a flowchart showing a flow of a denoising process in the third embodiment in which one type of learning model is included. In Figs. 14 to 16, the same parts as those in Figs. 2, 6, and 11 are denoted by the same reference numerals, and different parts will be described here.

In the third embodiment, the learning model of the denoising detector 2B is different from the learning model of the denoising detector 26 in the first and second embodiments, as shown in FIG. 15 or FIG. 16. FIG. 15 shows a processing flow in which a CNN learning model process (step S221), a difference output process (step S222), a learning model A (denoising model) (step S22A) including a shortcut in step S221, and a learning model B (step S22B) by a CNN learning model process (step S223) not including a shortcut are cascaded. In FIG. 15, the order is model A followed by model B, but the order may be reversed. Also, at least one of model A and model B is included.

The function of the shortcut is explained using Figure 16, which shows only model A. The shortcut is a line from the input of the CNN model to the CNN output, and is used to calculate the difference between the CNN output and the CNN input. As a result, when an unwanted signal N + target signal S is input as learning data and learning is performed so that the difference output becomes teacher data, the CNN model performs processing to simulate the unwanted signal, and S is extracted by subtracting the simulated N from the input data S + N. This makes it possible to suppress the unwanted signal and make it easier to detect the target signal.

Furthermore, by cascading model B, which does not include shortcuts, it is possible to carry out learning that extracts only the target signal, which is the teacher. In this case, the processing scale increases, but it becomes easier to extract only the target.

As described above, according to this embodiment, at least one of a learning model including shortcuts and a learning model not including shortcuts is used as the CNN model. In other words, when a CNN model using shortcuts and a model not including shortcuts are cascaded, the processing cost is the highest but the suppression performance is high, whereas when only one of them is used, the suppression performance is poor but the processing cost is low, allowing the optimal selection to be made according to the operation.

(Fourth embodiment) Sub-array DBF
In the first embodiment, a method for forming a Σ beam, a Σ AZ beam, and a Σ EL beam to perform squint angle measurement has been described. In this embodiment, a method for performing integrated processing by configuring both a phase monopulse beam and a squint beam with the same hardware using a subarray type DBF will be described with reference to Figs. 17 to 20.

Figure 17 shows the configuration of the receiving system of the radar system according to the fourth embodiment, Figure 18 shows an example in which the antenna aperture is divided into subarrays in the fourth embodiment, Figure 19 shows an example of the configuration of a subarray-type DBF applied to the fourth embodiment, and Figure 20 shows a coordinate system for explaining the phase for beam scanning of the received signals of the subarray in the fourth embodiment.

In the receiving system shown in Figure 17, the antenna aperture is divided into subarrays as shown in Figure 18(a) or 18(b), analog synthesis is performed within the subarray, and synthesis is performed by DBF between the subarrays. In the receiving system shown in Figure 17, subarrays 311 to 31n receive signals of Σ beams and ΣAZ beams squinted on the AZ axis, and Σ beams and ΣEL beams squinted on the EL axis, and the received signals are frequency converted by frequency converters 321 to 32n and converted to digital signals by AD converters 331 to 33n.

Then, the received output for each subarray is converted to the frequency domain by FFT/PC (Fast Fourier Transform/Pulse Compression) 341 to 34n through FFT calculation, and then DBF 35 performs DBF calculation on the received signals of the Σ beam, Σ AZ beam, and Σ EL beam in the specified direction, and the CFAR detector 36 and denoise detector 37 perform target detection, respectively. The integration processor 38 integrates the target detection results of each beam, and the observation value output unit 39 outputs the target observation value.

Meanwhile, the output of the integrated processor 38 is input sequentially to the cell extractors 401-40n, which compare the frequency signals for each subarray from the FFT/PC 341-34n with the processed signal from the integrated processor 38 to extract target detection cells, and the DBF monopulse beam former 41 and the AZ/EL phase monopulse beam former 42 obtain the amplitude and phase information of the target, respectively. The output of the cell extractor 40n is further input to the cell extractor 40n+1 to extract all cells in which the target exists, and the extraction results are sent to the DBF 35 to specify the direction of DBF formation, and the denoising angle measuring device 43 performs denoising angles for the Σ beam, Σ AZ beam, and Σ EL beam, and sends them to the integrated processor 44 together with the amplitude and phase information of the target from the AZ/EL phase monopulse beam former 42. The integration processor 44 integrates the amplitude, phase, and angle measurement values of each detected target, sends them as target information to the observation value output unit 39, and outputs them from the observation value output unit 39 together with the target detection results from the integration processor 38.

As shown in FIG. 19, the internal systems of each of the subarrays 311 to 31n perform low-noise amplification of the received signals of the antenna elements 511 to 514 using low-noise amplifiers 521 to 524, and the phase control for beam scanning is set by the receiving beam controller 55 using the phase shifters 531 to 534. The signals are then combined by the combiner 54 to obtain the subarray output.

Figure 10(a) shows the case where the aperture is divided into AZ4 x EL4, and Figure 10(b) shows the case where the aperture is divided into AZ2 x EL2, but it goes without saying that other division methods are also possible as long as the division is an even number. The phase within each subarray 311 to 31n is set so that a beam is formed in a specified AZ angle direction and EL angle direction, respectively. This makes it possible to form a sum beam Σ and a difference beam ΔAZ (ΔEL) by dividing the aperture into two, and phase monopulse angle measurement can be performed using equations (1) and (2).

On the other hand, the phase of each antenna element in the subarray remains the same, and a squint beam ΣAZ (ΣEL) can be formed by DBF between the subarrays.

To formulate this, the phase for beam scanning of the subarray signal is given by the following equation for the Σ beam, Σ AZ beam, and Σ EL beam, with reference to the coordinate system in Figure 20.

By using beams with these phase settings, highly accurate squint angle measurement can be achieved through the denoising processing of the first to third embodiments.

With the above processing, first, for detection, in the integration process (38), the result with the higher SNR is selected from the Σ beam before denoising processing and the Σ beam after denoising processing, and the target range-Doppler cell can be extracted. For angle measurement, in the integration process (44), if the SNR is higher before denoising processing, the phase monopulse is used, and in the opposite case, the squint angle measurement value after denoising processing is used to output the observation value. This makes it possible to reduce the variation in the detection results or angle measurement results due to denoising processing.

As described above, according to this embodiment, a squint beam is formed using a sub-array DBF, the SNR of the results of pre-denoising and post-denoising processing is compared for each target, and the larger range-Doppler cell is used to perform angle measurement processing using phase monopulse (squint angle measurement) in pre-denoising processing and squint angle measurement in post-denoising processing. In other words, by using a sub-array DBF, both the phase monopulse beam and the squint beam can be configured using the same hardware, and the detection or angle measurement processing of the pre-denoising and post-denoising processing can be integrated, thereby reducing the variation in the detection or angle measurement results due to the denoising processing.

Note that the above processing has been described as denoising processing, but it goes without saying that other processing such as super-resolution processing (SRCNN: Super-Resolution Convolution Neural Network, see Non-Patent Document 6) may also be applied.

In addition, the present invention is not limited to the above-described embodiment as it is, and in the implementation stage, the components can be modified and embodied without departing from the gist of the invention. Furthermore, various inventions can be formed by appropriately combining multiple components disclosed in the above-described embodiment. For example, some components may be deleted from all the components shown in the embodiment. Furthermore, components from different embodiments may be appropriately combined.

11: signal generator, 12: modulator, 13: frequency converter, 14: pulse modulator, 15: transmitting antenna,
21...receiving antenna, 22...frequency converter, 23...AD converter, 24...FFT/PC (Fast Fourier Transform/Pulse Compression) processor, 25...DBF (Digital Beam Forming) processor, 26...denoising detector, 27...observation value output unit, 28...cell extractor, 29...denoising angle measuring unit, 2A...pre-processor, 2B...denoising detector,
311 to 31n...subarray, 321 to 32n...frequency converter, 331 to 33n...AD converter, 341 to 34n...FFT/PC, 35...DBF, 36...CFAR detector, 37...denoise detector, 38...integration processor, 39...observation value output unit, 401 to 40n, 40n+1...cell extractor, 41...DBF monopulse beamformer, 42...AZ/EL phase monopulse beamformer, 43...denoise angle finder, 44...integration processor.

Claims (5)

A radar system that performs coherent integration processing on a slow-time axis using a single pulse or modulated N (N≧1) pulses,
a learning means for acquiring range-Doppler data in advance, dividing the range-Doppler data into M ways while overlapping on a range axis and a Doppler axis, and training a CNN (Convolutional Neural Network) model using learning data that combines the divided unnecessary wave data with previously acquired target data that serves as a teacher signal;
A radar system comprising: a processing means for dividing range-Doppler data of a Σ beam and a squint beam on the AZ (Azimuth) axis (EL (Elevation) axis) input in real time into M ways using a learning model based on the learning data, and using the range-Doppler data synthesized by denoising for each division unit, detecting targets by target detection processing using a CFAR (Constant False Alarm Rate) or maximum value using the Σ beam, and performing angle measurement processing using the Σ beam and the squint beam on the AZ axis (EL axis). The radar system of claim 1, wherein the learning means calculates a normalization coefficient for each divided data using the maximum amplitude of the Σ beam, and applies the same normalization coefficient to the Σ beam and the squint beam for processing. The radar system according to claim 1 or 2, wherein the learning means uses at least one of a learning model including a shortcut and a learning model not including a shortcut as the CNN model. The radar system according to any one of claims 1 to 3, wherein the processing means forms the squint beam by subarray type digital beamforming, compares the signal-to-noise ratio (SNR) for each target for the results of processing before denoising and processing after denoising, and performs angle measurement processing using the range-Doppler cell with the larger SNR, phase monopulse in processing before denoising, and squint angle measurement in processing after denoising. A radar signal processing method for performing coherent integration processing on a slow-time axis using a single pulse or modulated N (N≧1) pulses, comprising the steps of:
Obtain range-Doppler data in advance,
The range-Doppler data is divided into M ways while overlapping on the range axis and the Doppler axis;
A CNN (Convolutional Neural Network) model is trained using training data obtained by combining the divided unnecessary wave data and target data that is a teacher signal obtained in advance;
Using a learning model based on the learning data, the range-Doppler data of the Σ beam and the squint beam of the AZ (Azimuth) axis (EL (Elevation) axis) input in real time is divided into M ways,
Using the range-Doppler data that has been denoised and synthesised for each division unit, targets are detected using the Σ beam based on the CFAR (Constant False Alarm Rate) or maximum value.
A radar signal processing method for performing angle measurement processing using the Σ beam and a squint beam on the AZ axis (EL axis).

Images