Intelligent Target Recognition System and Method for Narrowband and Wideband Combined Sum-Difference Monopulse Radar

Abstract

The present application provides a recognition radar method and system based on combined narrowband and broadband sum-difference monopulse. The method includes: when the system is operating normally in the narrowband mode, it processes by transmitting a wide beam and simultaneously forming multiple high-gain narrow beams that cover the spatial domain of the transmitting beam; when it is necessary to identify a specific target, in the narrowband mode, a broadband mode time slot is inserted to form sum-difference triple beams, and the outputs of each beam are processed respectively; in the broadband mode, using the distance and velocity estimation results in the narrowband mode as prior information, adopting the sum-difference ratio amplitude monopulse angle measurement algorithm to obtain a high-resolution range-Doppler spectrogram of P×Q for the target and a strong scatterer distribution map of Y×Z; constructing a two-channel ResNet deep convolutional neural network, training the network parameters using the high-resolution range-Doppler spectrogram and the strong scatterer distribution map, and using the trained network for real-time target recognition. The present application is beneficial to target recognition.

Description

Technical Field

[0001] This application relates to the technical field of radar system target recognition, and particularly to an intelligent target recognition system and method based on combined narrow and wideband sum-difference monopulse radar. Background Art

[0002] Radar uses radio signals to detect and discover targets, and determine the position parameters of targets as well as their shape, structure, size, and motion characteristics, etc. It is an important way to obtain intelligence in modern warfare. With the increasingly complex battlefield electromagnetic environment and interference environment, the ability to quickly detect and identify targets has become a key link in the war.

[0003] Existing radar target recognition methods mainly use target high-resolution range profiles or micro-motion information. In terms of high-resolution range profile recognition, by using signals with a large instantaneous bandwidth, the range resolution can be improved to obtain the target's high-resolution one-dimensional range profile and the distribution of scattering points of the target along the radar line of sight. Therefore, it can be judged and recognized through the distribution characteristics of scattering points on the line of sight. This method is relatively effective for targets with large electrical sizes, such as aircraft and ships, etc., but for small targets such as vehicles, people, and animals, the recognition effect will be poor. Moreover, this method has the problem of being sensitive to the target's perspective. It is necessary to perform different analytical processes on the one-dimensional range profiles of the same target under different perspectives to obtain templates of a single target under multiple perspectives. When recognizing, it is necessary to match multiple groups of templates of different targets, with a large amount of calculation and low efficiency.

[0004] In addition, micro-motion refers to the tiny motion of the target itself or its components other than the centroid translation, such as vibration, rotation, and acceleration motion, etc. By modeling and analyzing the motion characteristics of the target, the differences in the micro-motion information of different targets are obtained, and this difference can be used to achieve target recognition. This method based on micro-motion characteristics recognition is applicable to situations where there are obvious differences in motion characteristics, such as wheeled vehicles and tracked vehicles, people and animals, etc., but does not consider the differences in target shape and contour.

[0005] Traditional target recognition methods based on micro-Doppler or micro-motion characteristics extract some physically meaningful features manually and use classifiers such as support vector machines and linear discriminant methods for classification and recognition. Such methods cannot extract the deep potential characteristics of the target and have the characteristics of a small amount of calculation and low recognition efficiency. Summary of the Invention

[0006] This application provides an intelligent target recognition system and method based on combined narrow and wideband sum-difference monopulse radar, which can be used to solve the technical problem of being unable to recognize deep potential characteristics during the target recognition process.

[0007] This application provides an intelligent target recognition method based on combined narrow and wideband sum-difference monopulse radar, and the method includes:

[0008] Step 1: The system normally operates in the narrowband mode. It conducts target search, confirmation, and range, velocity, and angle tracking by transmitting wide beams and simultaneously forming multiple high-gain narrow beams that cover the spatial domain of the transmitting beams.

[0009] Step 2: When it is necessary to identify a specific target, in the narrowband mode, insert broadband mode time slots. Using the angle estimation results in the narrowband mode, form sum and difference triple beams, and perform pulse compression and MTD processing on the output of each beam respectively.

[0010] Step 3: In the broadband mode, using the range and velocity estimation results in the narrowband mode as prior information, adopt the sum and difference amplitude comparison monopulse angle measurement algorithm to obtain a high-resolution range-Doppler spectrogram of P×Q for the target, and a strong scatterer distribution map of Y×Z.

[0011] Step 4: Construct a two-channel ResNet deep convolutional neural network, use the labeled high-resolution range-Doppler spectrogram and strong scatterer distribution image data to train the network parameters, and use the trained network for real-time target recognition.

[0012] Optionally, in the broadband mode, using the range and velocity estimation results in the narrowband mode as prior information, adopt the sum and difference amplitude comparison monopulse angle measurement algorithm to obtain a high-resolution range-Doppler spectrogram of P×Q for the target, and a strong scatterer distribution map of Y×Z, including:

[0013] Step 3.1: For the received broadband linear frequency modulation pulse echo signal, perform sum and difference triple beam formation, pulse compression, and MTD to obtain the range-Doppler spectrograms of the target sum beam, azimuth difference beam, and elevation difference beam.

[0014] Step 3.2: Set a decision threshold, perform threshold discrimination on the range-Doppler spectrogram of the sum beam, and determine the strong scatterers of the target that are greater than the threshold.

[0015] And calculate the range of the strong scatterers and the ratio of the outputs of the sum beam and azimuth difference beam within the range gate to obtain the azimuth angle distribution map of the scatterers.

[0016] Step 3.3: Describe the scatterer distribution using the scatterer angle and range distribution map.

[0017] Optionally, construct a two-channel ResNet deep convolutional neural network, use the labeled high-resolution range-Doppler spectrogram and strong scatterer distribution image data to train the network parameters, and use the trained network for real-time target recognition, including:

[0018] Step 4.1: Divide the obtained range-Doppler spectrogram and strong scatterer distribution map into a training set and a validation set, and make corresponding labels.

[0019] Step 4.2: Set training parameters including the loss function, learning rate, and maximum number of iterations. Put the training set and labels into the improved dual-channel ResNet deep convolutional neural network, and perform parameter training until the training loss converges, then save the obtained training parameters.

[0020] Step 4.3: Load the trained weight parameters into the dual-channel ResNet deep convolutional neural network to obtain the final network model.

[0021] Optionally, in the dual-channel ResNet deep convolutional neural network, one channel is used to input the scatter point angle map, and the other channel is used to input the distance distribution map.

[0022] The two channels have the same structure. Each channel includes 4 convolutional layers and a pooling layer, followed by 3 fully connected layers. The two channels are connected at the fully connected layer and share the subsequent 2 common fully connected layers.

[0023] This application also provides an intelligent target recognition system based on the combination of wideband and narrowband sum-difference monopulse radar. The system includes:

[0024] Subarray phased array digital transceiver antennas, 2N-channel RF channel receivers, and 2N-channel wideband high-speed sampling / processing / recognition subsystems;

[0025] Among them, every M antennas out of 2M×N unit antennas form a subarray output through the T / R component and power combining network. The 2N-channel RF channel receivers mix and filter the RF signals and then output 2N-channel intermediate frequency signals, which are output to the 2N-channel wideband high-speed sampling / processing / recognition subsystems.

[0026] Optionally, the 2N-channel wideband high-speed sampling / processing / recognition subsystems are integrated in a chassis that conforms to the 6U VPX structure of the VITA46 standard.

[0027] The board card modules in the 2N-channel wideband high-speed sampling / processing / recognition subsystems include a multi-channel sampling module, a V7XC7VX690T FPGA and DSP processing module, a TMS320C6678 DSP processing module, and a Jetson AGX Xavier GPU processing module.

[0028] Among them, the FPGA is used for parallel processing of multiply-accumulate operations.

[0029] The DSP processing module is used for fast operation of complex signal processing algorithms.

[0030] The GPU with hundreds of CUDA cores is used to build a deep neural network.

[0031] Between the board modules, there is an SRIO switching network and a point-to-point high-speed star interconnect network, and the bi-directional interaction bandwidth between any two boards is not less than 40 Gbps.

[0032] Compared with the prior art, this application:

[0033] The radar system is built on a hardware platform that complies with the VITA466U standard, and has the characteristics of standardization, modularity, high speed and high performance. The board modules based on heterogeneous processing nodes such as FPGA, multi-core DSP and GPU can be configured as needed to give full play to the advantages of various processing nodes. Between the board modules, there is an SRIO switching network and a point-to-point high-speed star interconnect network, which meets the data rate brought by large bandwidth;

[0034] The narrowband mode and the broadband mode are dynamically configured for use as needed. The system normally operates in the narrowband mode. When it is necessary to identify specific targets, a broadband mode with an accumulation period is inserted, and it is possible to detect and identify at the same time.

[0035] The high-resolution range-Doppler spectrogram and the strong scatterer distribution map of the target contain rich characteristic information of complex targets. Among them, it not only contains the distribution and micro-motion information of the scatterers on the radar line of sight of the target, etc., but also contains information such as the rough outline of the target, which is more conducive to target recognition.

[0036] An improved dual-channel ResNet deep convolutional neural network is constructed for feature extraction and fusion of the range-Doppler spectrogram and the scatterer distribution map, realizing high-precision target recognition, and the recognition performance is greatly improved compared with the existing single-channel network.

[0037] Use the GPU development board platform to deploy the dual-channel convolutional neural network, and utilize the parallel processing ability of the GPU for large-scale data to meet the requirements of real-time processing. In addition, deploying the convolutional neural network on the GPU platform is simple and convenient, and has the advantages of low cost and good flexibility. Brief Description of the Drawings

[0038] Figure 1 It is the flowchart of the method provided by the embodiment of this application;

[0039] Figure 2 It is the schematic diagram of the phased array sub-array digital array antenna structure provided by the embodiment of this application;

[0040] Figure 3 It is the example diagram of the external shape structures of four typical wheeled and tracked vehicles provided by the embodiment of this application;

[0041] Among them, 3(a) is one of the wheeled vehicles, 3(b) is the second wheeled vehicle, 3(c) is one of the tracked vehicles, and 3(d) is the second tracked vehicle;

[0042] Figure 4High-resolution range-Doppler spectra of four typical wheeled and tracked vehicles provided in the embodiments of this application;

[0043] Among them, 4(a) is one of the wheeled vehicles, 4(b) is two of the wheeled vehicles, 4(c) is one of the tracked vehicles, and 4(d) is two of the tracked vehicles;

[0044] Figure 5 Distribution diagram of strong scattering points for four typical wheeled and tracked vehicles provided in the embodiments of this application;

[0045] Among them, 5(a) is one of the wheeled vehicles, 5(b) is two of the wheeled vehicles, 5(c) is one of the tracked vehicles, and 5(d) is two of the tracked vehicles;

[0046] Figure 6 A structural diagram of an improved dual-channel convolutional neural network used in the system provided in an embodiment of the present application;

[0047] Figure 7 The loss and accuracy curves during the training process provided in the embodiment of this application. DETAILED DESCRIPTION

[0048] In order to make the objectives, technical solutions and advantages of the present application clearer, the implementation methods of the present application will be further described in detail below with reference to the accompanying drawings.

[0049] The following first introduces the embodiments of the present application with reference to the accompanying drawings.

[0050] This application provides an intelligent target recognition method based on wide- and narrow-band combined and differential monopulse radar, the method comprising:

[0051] Step 1: The system operates normally in narrowband mode, performing target search, confirmation, and range, velocity, and angle tracking by transmitting a wide beam and simultaneously forming multiple high-gain narrow beams covering the transmit beam airspace.

[0052] Step 2: When a specific target needs to be identified, in narrowband mode, insert the broadband mode time slot and use the angle estimation results in narrowband mode to form three beams of sum and difference. Pulse compression and MTD processing are performed on the output of each beam.

[0053] Step 3: In broadband mode, the distance and velocity estimation results in narrowband mode are used as prior information, and the sum-difference amplitude ratio single pulse angle measurement algorithm is adopted to obtain the target's P×Q high-resolution range-Doppler spectrum and Y×Z strong scattering point distribution map.

[0054] Specifically, in step 3.1: For the received broadband linear frequency modulation pulse echo signal, after sum-difference three-beamforming, pulse compression, and MTD, the range-Doppler spectrograms of the target sum beam, azimuth difference beam, and elevation difference beam are obtained;

[0055] In step 3.2: Set the decision threshold, perform threshold discrimination on the range-Doppler spectrogram of the sum beam, and those greater than the threshold are determined as strong scattering points of the target;

[0056] And calculate the ratio of the output of the sum beam and the azimuth difference beam within the range gate of the distance of the strong scattering point to obtain the azimuth angle distribution map of the scattering point;

[0057] In step 3.3: Since the distance from the target to the radar is relatively far, the angles of multiple scattering points of the target are concentrated. In order to better represent the positional relationship between the scattering points, the scattering point distribution is described using the scattering point angle and distance distribution map, that is, the target strong scattering point distribution map.

[0058] In step 4: Construct a two-channel ResNet deep convolutional neural network, use the high-resolution range-Doppler spectrogram and the strong scattering point distribution image data to train the network parameters, and use the trained network for real-time target recognition.

[0059] Specifically, in step 4.1: Divide the obtained range-Doppler spectrogram and the strong scattering point distribution map into a training set and a validation set, and make corresponding labels;

[0060] In step 4.2: Set training parameters including the loss function, learning rate, and maximum number of iterations. Put the training set and labels into the improved two-channel ResNet deep convolutional neural network for parameter training until the training loss converges, and then save the obtained training parameters;

[0061] In step 4.3: Load the trained weight parameters into the two-channel ResNet deep convolutional neural network to obtain the final network model for high-precision target recognition. One channel in the two-channel ResNet deep convolutional neural network is used to input the high-resolution range-Doppler spectrogram, and the other channel is used for the strong scattering point distribution map;

[0062] The two-channel structures are the same. Each channel includes 4 convolutional layers and a pooling layer, followed by 3 fully connected layers. The two channels are connected at the fully connected layer and share the subsequent 2 common fully connected layers.

[0063] This application also provides an intelligent target recognition system based on a combined narrowband and broadband sum-difference monopulse radar. The system includes:

[0064] Subarray phased array digital transceiver antenna, 2N-channel RF channel receiver, and 2N-channel broadband high-speed sampling / processing / recognition sub-system.

[0065] Among them, every M antennas out of the 2M×N unit antennas form a sub-array output through the T / R components and the power combining network. The 2N-channel receiver mixes and filters the RF signals and then outputs 2N intermediate-frequency signals, which are output to the 2N-channel wideband high-speed sampling / processing / identification subsystem.

[0066] The 2N-channel wideband high-speed sampling / processing / identification subsystem is integrated in a chassis that conforms to the 6U VPX structure of the VITA46 standard;

[0067] The board modules in the 2N-channel wideband high-speed sampling / processing / identification subsystem include a multi-channel sampling module, a V7XC7VX690T FPGA and DSP processing module, a TMS320C6678 DSP processing module, and a Jetson AGX Xavier GPU processing module;

[0068] Among them, the FPGA with a large number of logic and timing resources is used for parallel processing of simple multiply-accumulate operations;

[0069] The multi-core DSP processing module is used for fast operation of complex signal processing algorithms;

[0070] The GPU with hundreds of CUDA cores is used to conveniently construct complex and flexible deep neural networks;

[0071] The board modules are interconnected through an SRIO switching network and a point-to-point high-speed star interconnection network, and the two-way interaction bandwidth between any two boards is not less than 40 Gbps.

[0072] The following specifically describes the present invention in detail with reference to the accompanying drawings.

[0073] Embodiment

[0074] The present invention is a radar system for fast detection and intelligent identification of complex targets. The system normally operates in the narrowband mode. When specific target identification is required, a broadband mode with an accumulation period is inserted, and detection and identification can be achieved simultaneously. The method flow is as Figure 1 shown, including three key steps: detecting in the narrowband mode and determining the position, speed, and angle parameters of the target; obtaining the high-resolution range-Doppler spectrogram and the strong scatterer distribution map of the target in the broadband mode; using an improved two-channel ResNet deep convolutional neural network to achieve high-precision identification.

[0075] The array antenna in this embodiment adopts a rectangular array structure of 48×24 units, and the structure diagram is as Figure 2 shown. In the figure, d1 = 9.8 mm and d2 = 13.5 mm. The four typical wheeled and tracked vehicles to be identified are shown in the external structure diagram as Figure 3 shown. The signal waveform used by the system is an LFM signal. The specific parameter settings are as follows: pulse width τ = 60 us, pulse repetition period Tr = 360 us, the pulse accumulation period N_PRI = 128, and the sampling frequency f of the system for the intermediate frequency signals of 48 channels s = 650 MHz. The narrowband bandwidth is 30 MHz, and the wideband bandwidth is 250 MHz.

[0076] The method for detection and recognition includes the following steps:

[0077] Step 1: The system normally operates in the narrowband mode, emits a wide beam covering the pitch angle of ±7.5°, and conducts target search, confirmation, and range, velocity, and angle tracking by simultaneously forming 9 high-gain narrow beams covering the airspace of the transmitting beam.

[0078] Step 2: When it is necessary to identify a specific target, in the narrowband mode, insert a wideband mode time slot, use the angle estimation result in the narrowband mode, form sum and difference three beams at this angle, and perform pulse compression and MTD processing on the output of each beam respectively.

[0079] Step 3: In the wideband mode, use the range and velocity estimation results in the narrowband mode as prior information, adopt the sum and difference monopulse angle measurement algorithm to obtain a high-resolution range-Doppler spectrogram of 256×128 for the target, and a strong scatterer distribution map of 140×140. The specific steps are as follows:

[0080] Step 3.1: Perform sum and difference three-beam formation, pulse compression, and MTD on the wideband linear frequency modulation pulse echo signal after being reflected by the target to obtain the range-Doppler spectrograms of the target sum beam, azimuth difference beam, and pitch difference beam.

[0081] Step 3.2: Considering that the maximum length of the target to be identified does not exceed 12 m, in the range dimension, the range corresponding to one sampling point is 0.23 meters. Therefore, the system selects 256 sampling points to represent the distribution of the target in the range dimension. After the echo signal passes through pulse compression and MTD, a range-Doppler spectrogram of 128×256 for the sum beam is obtained. This image contains the radial distribution characteristics, velocity characteristics, and micro-motion characteristics of the scatterers of the target in the radar line-of-sight direction, and is one of the image data input to the neural network. Figure 4 It is the high-resolution range-Doppler spectrogram obtained when 4 types of targets are at azimuth and pitch angles of 90° and 0° respectively, and the velocity is 4 m / s.

[0082] Step 3.3: Set the decision threshold to 0.3 times the maximum peak value of the range-Doppler spectrogram, and perform threshold discrimination on the peak value of the sum beam range-Doppler spectrum. If the peak point is greater than the threshold, it is determined as a strong scatterer. At the same time, calculate the range of this scatterer and the ratio of the outputs of the sum beam and azimuth difference beam within this range gate to obtain the azimuth angle of the scatterer;

[0083] Step 3.4: Since the distance from the target to the radar is relatively far, the azimuth angles of the scattering points of the target differ slightly. Extract the scattering points in the target area. To better represent the positional relationship between the scattering points, use the distance and azimuth angle information of the scattering points and convert them into points in the rectangular coordinate system using formula (1). Among them, ρ i is the distance of the i-th strong scattering point, θ i is the azimuth angle of the i-th strong scattering point, (x0, y0) is the coordinate value closest to the radar in the extracted target area, and (x i , y i ) is the rectangular coordinate value of the i-th strong scattering point. The two-dimensional map obtained through the above processing is the strong scattering point distribution map of the target. Given that the maximum length of the vehicle is less than 12 meters, the representation result is converted into an image form at intervals of 0.1 meter, with a size of 140×140. This is one of the image data input into the neural network and contains rough contour, shape and other features of the vehicle. Figure 5 The strong scattering point distribution maps obtained when the azimuth angle and elevation angle are 90° and 0° respectively, and the speed is 4 m / s for 4 kinds of targets are shown as follows.

[0084]

[0085] Step 4: Construct an improved dual-channel ResNet deep convolutional neural network for feature extraction and fusion of the range-Doppler spectrogram and the scattering point distribution map to achieve high-precision target recognition. Use the above labeled image data to train the network parameters, and use the trained network for real-time target recognition. The specific steps are as follows:

[0086] Step 4.1: Divide the obtained target range-Doppler spectrogram and strong scattering point distribution map into a training set and a validation set, and make corresponding labels.

[0087] Step 4.2: Put the training set and labels into the dual-channel ResNet deep convolutional neural network for training. Among them, there are multiple convolutional pooling layers in each channel for extracting the features of the target. At the same time, considering that the "image" to be recognized is relatively simple, to prevent overfitting in training, the depth of the convolutional pooling layer should not be too deep. The fully connected layers of the two channels are used for feature fusion of the two images. The improved ResNet dual-channel network model is shown as Figure 6 follows.

[0088] Use the Adam optimization algorithm to optimize the network model parameters. During training, the cross-entropy function is used as the loss function, the learning rate is set to 0.0001, the maximum number of iterations is 15 times, and the learned parameters are saved. Figure 7 As can be seen from the loss and accuracy curve change diagram with the increase of the training iteration times, as the training iteration progresses, the loss becomes smaller and the accuracy becomes higher, gradually converging and stabilizing.

[0089] Step 4.3: Load the trained weight parameters into the designed neural network to obtain the final network model, thereby achieving high-precision recognition. The final recognition rates are shown in the following table. It can be seen that the recognition rates are all above 89%, with an average of 96.4%.

[0090] Table 1: Recognition Rate Table

[0091]

[0092] Step 5: The above narrowband mode and broadband mode are configured and used as needed. The system normally operates in the narrowband mode. When specific target recognition is required, insert the broadband mode for one accumulation period, and the effect of detecting and recognizing simultaneously can be achieved.

[0093] Compared with the prior art, the present application:

[0094] The radar system is constructed using a hardware platform that complies with the VITA466U standard, featuring standardization, modularity, high speed, and high performance. The board modules based on heterogeneous processing nodes such as FPGA, multi-core DSP, and GPU can be configured as needed to fully utilize the advantages of various processing nodes. The board modules are interconnected through the SRIO switching network and the point-to-point high-speed star interconnection network, meeting the data rate requirements brought by large bandwidth;

[0095] The narrowband mode and broadband mode are dynamically configured and used as needed. The system normally operates in the narrowband mode. When specific target recognition is required, insert the broadband mode for one accumulation period, and the effect of detecting and recognizing simultaneously can be achieved.

[0096] The high-resolution range-Doppler spectrogram and strong scatterer distribution map of the target contain rich characteristic information of complex targets. Among them, it not only contains the distribution and micro-motion information of the scatterers on the radar line of sight of the target, but also contains information such as the rough outline of the target, which is more conducive to target recognition.

[0097] An improved dual-channel ResNet deep convolutional neural network is constructed for feature extraction and fusion of the range-Doppler spectrogram and scatterer distribution map, achieving high-precision target recognition, and greatly improving the recognition performance compared with the existing single-channel network.

[0098] Deploy the dual-channel convolutional neural network using the GPU development board platform. Utilize the parallel processing ability of the GPU for large-scale data to meet the requirements of real-time processing. In addition, deploying the convolutional neural network on the GPU platform is simple and convenient, with the advantages of low cost and good flexibility.

[0099] The above-described embodiments of the present application do not constitute a limitation on the protection scope of the present application.

Claims

1. A method for intelligent target recognition based on a combined narrowband and broadband sum-difference monopulse radar, characterized in that, The method includes: Step 1: When the system is operating normally in the narrowband mode, target search, confirmation, and distance, speed, and angle tracking are performed by transmitting a wide beam and simultaneously forming multiple high-gain narrow beams that cover the airspace of the transmitting beam. Step 2: When it is necessary to identify a specific target, in the narrowband mode, a broadband mode time slot is inserted. Using the angle estimation results in the narrowband mode, sum-difference triple beams are formed, and pulse compression and MTD processing are performed on the outputs of each beam respectively. Step 3: In the broadband mode, using the distance and speed estimation results in the narrowband mode as prior information, the sum-difference amplitude-comparison monopulse angle measurement algorithm is adopted to obtain a high-resolution range-Doppler spectrogram of P×Q for the target and a strong scatterer distribution map of Y×Z. Step 3.1: For the received broadband linear frequency modulation pulse echo signal, after sum-difference triple beam formation, pulse compression, and MTD, range-Doppler spectrograms of the target sum beam, azimuth difference beam, and elevation difference beam are obtained. Step 3.2: A decision threshold is set, and the range-Doppler spectrogram of the sum beam is subjected to threshold discrimination. Those greater than the threshold are determined to be strong scatterers of the target. And the distance of the strong scatterer and the ratio of the outputs of the sum beam and the azimuth difference beam within the distance gate are calculated to obtain the azimuth angle distribution map of the scatterer. Step 3.3: The scatterer distribution is described using the scatterer angle and distance distribution map. Step 4: A two-channel ResNet deep convolutional neural network is constructed. The network parameters are trained using the labeled high-resolution range-Doppler spectrogram and strong scatterer distribution image data, and the trained network is used for real-time target recognition. Step 4.1: The obtained range-Doppler spectrogram and strong scatterer distribution map are divided into a training set and a validation set, and corresponding labels are made. Step 4.2: Training parameters including a loss function, a learning rate, and a maximum number of iterations are set. The training set and labels are put into the improved two-channel ResNet deep convolutional neural network for parameter training until the training loss converges, and then the obtained training parameters are saved. Step 4.3: The trained weight parameters are loaded into the two-channel ResNet deep convolutional neural network to obtain the final network model.

2. The intelligent target recognition method based on the combination of wide and narrow bands and sum-difference monopulse radar according to claim 1, characterized in that One channel in the two-channel ResNet deep convolutional neural network is used for the high-resolution range-Doppler spectrogram, and the other channel is used for the strong scatterer distribution map. The two-channel structures are the same. Each channel includes 4 convolutional layers and a pooling layer, followed by 4 fully connected layers. The two channels are connected at the fully connected layer and share the subsequent 2 common fully connected layers.

3. A narrowband and broadband combined sum-difference monopulse radar intelligent target recognition system for implementing any one of the methods in claims 1 to 2, characterized in that, It includes: A subarray phased array digital transceiver antenna, a 2N-channel RF channel receiver, and a 2N-channel broadband high-speed sampling / processing / recognition subsystem; Among them, every M antennas out of 2M×N unit antennas form a subarray output through a T / R component and a power combining network. The 2N-channel receiver mixes and filters the RF signals and outputs 2N intermediate frequency signals, which are then output to the 2N-channel broadband high-speed sampling / processing / recognition subsystem.

4. The intelligent target recognition system based on narrowband and wideband combined sum-difference monopulse radar according to claim 3, wherein The 2N-channel broadband high-speed sampling / processing / recognition subsystem is integrated in a chassis that conforms to the 6U VPX structure of the VITA46 standard; The board card module in the 2N-channel broadband high-speed sampling / processing / recognition subsystem includes a multi-channel sampling module, a V7XC7VX690T FPGA and DSP processing module, a TMS320C6678 DSP processing module, and a Jetson AGX Xavier GPU processing module; Among them, the FPGA is used for parallel processing of multiply-accumulate operations; the DSP processing module is used for fast operation of complex signal processing algorithms; The GPU with hundreds of CUDA cores is used to build a deep neural network; The board card modules are interconnected through an SRIO switching network and a point-to-point high-speed star interconnection network, and the two-way interaction bandwidth between any two board cards is not less than 40 Gbps.

Images