A signal processing method for multi-dimensional radar target intelligent identification

Abstract

This invention discloses a signal processing method for intelligent multi-dimensional radar target recognition, relating to the technical field of information perception and recognition technology. Unlike traditional radar target classification and recognition methods, this method employs various modulated waveform transmission modes, including narrowband, wideband, and long-term accumulation, and utilizes multiple signal processing modes such as general spectrum analysis, high-resolution range processing, time-frequency processing, multi-dimensional spectrum analysis, multi-frame correlation processing, and time-range-Doppler transform. This yields target feature results such as range, velocity, energy, RCS, one-dimensional range profile, micro-motion Doppler, time-frequency features, motion trends, and multi-dimensional information. Through deep learning-based multi-mode target feature extraction, combined with target recognition criteria, it achieves comprehensive multi-mode target perception and recognition. Compared to traditional target recognition and AI recognition methods, this method features high recognition rate, simple feature extraction, and strong scalability.

Description

Technical Field

[0001] This invention relates to the technical field of information perception and recognition technology, and belongs to the innovative type of multi-dimensional radar target intelligent recognition signal processing method, which is applicable to target detection and recognition applications in reconnaissance and detection radar systems. Background Technology

[0002] Currently, traditional radar target recognition, after target detection, uses target recognition feature extraction algorithms to calculate target feature values ​​based on experience to achieve target classification and recognition. Intelligent recognition methods, based on an established target feature database, employ deep learning techniques to simultaneously perform target detection and recognition, achieving intelligent target detection and recognition. However, traditional methods only use single-dimensional target signal features for target recognition, resulting in low recognition accuracy. Therefore, there is an urgent need for a multi-dimensional radar target intelligent recognition method that employs multi-dimensional signal processing algorithms to conduct deep learning-based target detection and recognition, improving the reliability and feasibility of intelligent recognition. Summary of the Invention

[0003] This invention discloses a signal processing method for multi-dimensional radar target intelligent recognition, involving radar transmitted waveform modulation, received signal demodulation, multi-mode signal processing, deep learning target feature extraction, and multi-mode recognition, applied to radar detection, perception, and recognition. Unlike traditional radar target classification and recognition methods, this method employs narrowband, wideband, and long-term accumulation modulated waveform transmission, along with general spectrum analysis, high-resolution range processing, time-frequency processing, multi-dimensional spectrum analysis, multi-frame correlation processing, and time-range-Doppler transform, among other signal processing methods, to obtain corresponding target feature results such as range, velocity, energy, RCS, one-dimensional range profile, micro-motion Doppler, time-frequency features, motion trends, and multi-dimensional information. Through deep learning multi-mode target feature extraction, combined with target recognition criteria, it achieves comprehensive multi-mode target perception and recognition. Compared to traditional target recognition and AI recognition methods, this method features high recognition rate, simple feature extraction, and strong scalability.

[0004] The technical solution adopted in this invention is as follows:

[0005] A signal processing method for intelligent multidimensional radar target recognition includes the following steps:

[0006] Step 1: Radar transmit waveform modulation. The radar transmit waveform uses multiple modulation methods to achieve the transmission of multi-modulated signals, including narrowband modulation, wideband modulation, and long-time frame-accumulation modulation. Among them, narrowband modulation includes frequency modulation, phase modulation, and pulse train modulation; wideband modulation includes frequency modulation, phase modulation, pulse train modulation, and composite modulation; long-time frame-accumulation modulation includes step frequency modulation and space-time-frequency coded composite modulation.

[0007] Step 2, Received Signal Demodulation: The radar demodulates the received signal according to the modulation scheme of the transmitted signal. The echo signal generated after the radar transmits the modulated signal and encounters the target is received by the antenna, amplified by radio frequency, frequency converted, demodulated, and converted from analog to digital to form the received signal. The demodulation transformation is performed on the received signal after long-term accumulation, code correlation, frequency modulation demodulation, and phase demodulation to form a signal processing and acquisition digital signal. All targets are known targets, and the target category is recorded as the truth label of the target.

[0008] Step 3, multi-mode signal processing, through spectrum analysis, time-frequency analysis, range correlation demodulation, long-term accumulation spectrum analysis, high-resolution range demodulation, multi-dimensional spectrum analysis, and time-range-Doppler transform processing, to obtain the multi-dimensional information feature results of the target;

[0009] Step 4: Deep learning target feature extraction. Deep learning network training is carried out using convolutional neural networks to target multidimensional information feature results. The network layer parameters are associated with the target features, and the trained convolutional neural network model is saved.

[0010] Step 5, multi-mode recognition: Through parallel / time-division multi-mode signal processing, the transmitted waveform of the radar signal is modulated, the received signal is demodulated, and the multi-dimensional information feature results of the target are obtained. The multi-dimensional information feature results of the target are input into the trained convolutional neural network model to obtain the target classification result, thus completing the signal processing for multi-dimensional radar target intelligent recognition.

[0011] Furthermore, the multi-modulation signals in step 1 include linear frequency modulated radar signals, nonlinear frequency modulated radar signals, polynomial phase modulated radar signals, sinusoidal frequency modulated radar signals, phase shift keying radar signals, binary phase coded and quadrature phase coded radar signals, frequency coded radar signals, linear frequency modulated-phase coded radar signals, frequency coded-phase coded radar signals, step frequency modulated signals, and space-time frequency coded signals.

[0012] Furthermore, the specific method for step 3 is as follows:

[0013] Step 301: Spectrum analysis. Using conventional FFT transformation, the velocity, energy, and micro-doppler information of the target are obtained under narrowband modulation signal.

[0014] Step 302: Time-frequency analysis. Short-time FFT, WVD, and wavelet transform are used to obtain the time-frequency characteristic information of the target under narrowband modulation signal.

[0015] Step 303: Range-related demodulation, including frequency modulation demodulation and phase modulation demodulation, to obtain the target's range information parameters under narrowband modulation signal; combined with the energy information characteristics of the spectrum analysis results, to calculate the corresponding range echo energy and characterize the target's radar cross-section.

[0016] Step 304: Long-term accumulated spectrum analysis mainly relies on multi-frame continuous processing. It processes the long-term accumulated frame modulation signal to obtain the time-frequency transformation characteristics, micro-doppler, and motion trend characteristics of the target over a long period of time.

[0017] Step 305: High-resolution range demodulation. By using the range demodulation results of down-modulation and phase-modulation demodulation of the broadband modulated signal, high-resolution one-dimensional range image information and high-precision range change feature information of the target are obtained.

[0018] Step 306: Multidimensional spectrum analysis. For long-term accumulated processing, multidimensional spectrum analysis calculation is performed to obtain the target micro-motion Doppler information and target motion trend characteristic information.

[0019] Step 307: Time-range-Doppler transformation, for long-term accumulation processing and signal composite processing, to obtain the target's time, energy, distance, and Doppler multidimensional composite information.

[0020] Furthermore, step 4 is specifically implemented as follows:

[0021] Step 401: Input data, with a dimension of batch_size × input_dim; batch_size is the amount of data used in the training iteration, and input_dim is the dimension of the input layer data, that is, the dimension of the multidimensional information feature result of the target.

[0022] Step 402: Construct a convolutional neural network architecture, including an input layer, a hidden layer, and an output layer. The input layer has a data dimension of input_dim and an output dimension of output_dim. The hidden layer has an input dimension of output_dim and an output dimension of num_feature. The output layer has an input dimension of num_feature and an output dimension of num_classes.

[0023] Step 403: Forward propagation. The input data is fed into the convolutional neural network, passing through the input layer, hidden layer, and output layer, and outputting the target classification result. ;

[0024] Step 404: Calculate the loss and convert the positive output target classification result into a value. Truth labels corresponding to input data The two are compared, and the cross-entropy loss function is used to calculate the loss function value between them. ;

[0025] Step 405: Backpropagation, applying the calculated loss function value to the parameters in the convolutional neural network. Optimize;

[0026] Step 406: Repeat steps 403-405. When the loss function value no longer decreases for several consecutive rounds, stop the loop and obtain the final convolutional neural network model.

[0027] The advantages of this invention compared to the prior art are:

[0028] 1. This invention uses deep learning methods to extract target features, forming an integrated model of target detection and classification recognition. Feature extraction is simple and improves the flexibility of the system.

[0029] 2. This invention employs multi-mode signal processing, utilizing time-division processing and parallel processing, to acquire multiple features of the target, thereby improving the reliability of target recognition in the system;

[0030] 3. This invention employs multiple modes for simultaneous recognition, resulting in strong system scalability and improved target recognition probability. Attached Figure Description

[0031] Figure 1 This is a schematic flowchart of a signal processing method for intelligent recognition of multi-dimensional radar targets in an embodiment of the present invention. Detailed Implementation

[0032] The present invention will be further explained and described below with reference to the accompanying drawings and specific embodiments.

[0033] A signal processing method for intelligent target recognition using multi-dimensional radar is presented. A practical example of a radar employing this method utilizes narrowband, wideband, and long-term accumulation modulation waveforms for transmission. Multiple signal processing modes, including general spectrum analysis, high-resolution range processing, time-frequency processing, multi-dimensional spectrum analysis, multi-frame correlation processing, and time-range-Doppler transform, are employed to obtain target feature results such as range, velocity, energy, RCS, one-dimensional range profile, micro-motion Doppler, time-frequency features, motion trends, and multi-dimensional information. Through deep learning-based multi-mode target feature extraction and combined with target recognition criteria, comprehensive multi-mode target perception and recognition are achieved. Figure 1 As shown, the method includes the following steps:

[0034] (1) Practical example of radar transmit waveform modulation: Radar transmit waveforms are implemented using time division multiplexing, frequency division multiplexing, and code division multiplexing to achieve the transmission of various modulated signals. These include narrowband modulation waveforms such as frequency modulation (LFM, NLFM, SFM, FSK), phase modulation (PPS, PSK, BPSK, QPSK), and pulse train modulation; wideband modulation such as frequency modulation, phase modulation, composite pulse group modulation, and composite modulation (LFM-BPSK, FSK-BPSK); and long-time frame-accumulating modulation such as stepped frequency modulation (SFS) and space-time frequency coded modulation (STFBC). Simple narrowband and wideband transmit signals are generally implemented using time division multiplexing, frequency modulation narrowband and wideband transmit signals are generally implemented using frequency division multiplexing, phase modulation narrowband and wideband transmit signals are generally implemented using code division multiplexing, and composite modulation and long-time frame-accumulating modulation transmit signals can be implemented using frequency division and code division composite multiplexing.

[0035] (2) Practical example of received signal demodulation: The echo signal generated after the radar transmits a modulated signal encounters a target is received by the antenna, amplified by radio frequency, frequency converted, demodulated, and converted from analog to digital to form a received signal. For echo received signals with various waveforms, a digital / analog multiplexing method is used to realize the demodulation and transformation of the received signal. This includes signal integration processing, long-term accumulation processing of multi-frame received signals, digital / analog frequency modulation demodulation processing, and digital / analog phase decoding correlation demodulation processing.

[0036] (3) Practical example of multi-mode signal processing: For echoes of different modulated signals, multi-mode signal processing is carried out in parallel, including spectrum analysis using conventional FFT transform to obtain narrowband information such as target range, velocity, and angle; time-frequency analysis using short-time FFT, WVD, wavelet transform, etc. to obtain target time-frequency two-dimensional feature information; range-correlation demodulation using frequency modulation demodulation, phase modulation demodulation, etc. to obtain target high-resolution one-dimensional range image information and high-precision range change feature information; using multi-frame continuous long-term accumulated data, time-frequency transformation feature information, micro-motion Doppler information, motion trend information, etc. are obtained through spectrum analysis; micro-motion Doppler information, target motion trend feature information, multi-dimensional motion information, etc. can be obtained through multi-dimensional spectrum analysis; through time-range-Doppler transformation, relying on composite processing of signals over a long period of time, multi-dimensional super-resolution information such as target time, energy, range, and Doppler can be obtained; after calibration calculation for radar equation, the target radar cross-section RCS information can also be obtained, as well as multi-dimensional feature information of the target obtained through multi-dimensional signal processing transformation, etc. In practical applications, the multiplexed digital echo information obtained by demodulating the received signal can be used to calculate the results through processors such as DSPs, GPUs, and CPUs.

[0037] (4) Deep learning target feature extraction: Deep learning network training is carried out through convolutional neural network for the multidimensional information feature results of the target. The network layer parameters are associated with the target features, and the trained convolutional neural network model is saved.

[0038] The specific method for constructing deep learning networks is as follows:

[0039] Step 1: Input data, with dimensions of batch_size × input_dim; batch_size is the amount of data used in the training iteration, and input_dim is the dimension of the input layer data, i.e., the dimension of the multidimensional information feature results of the target.

[0040] Step 2: Construct the network architecture, including an input layer, hidden layers, and an output layer. The input layer has a data dimension of input_dim and an output dimension of output_dim. The hidden layer has an input dimension of output_dim and an output dimension of num_feature. The output layer has an input dimension of num_feature and an output dimension of num_classes, where num_classes=1. The output layer outputs the classification results of the target, such as the classification of targets like cars, houses, drones, trees, and people.

[0041] Step 3: Forward propagation, the data is input into the deep learning network, passing through the input layer, hidden layer, and output layer respectively, and the classification result is output. ;

[0042] Step 4: Calculate the loss and convert the positive output target classification result into a value. Truth labels corresponding to input data The two are compared, and the cross-entropy loss function is used to calculate the loss function value between them. ;

[0043] Step 5: Backpropagation, applying the calculated loss function value to the parameters in the deep learning network model. Optimize;

[0044] Step 6: Repeat steps 3-5. When the loss function value no longer decreases for several consecutive rounds, stop the loop and obtain the final classification deep learning network.

[0045] (5) Multi-mode recognition: Through parallel / time-division multi-mode signal processing, the transmitted waveform of the radar signal is modulated, the received signal is demodulated, and the multi-dimensional information feature results of the target are obtained. The multi-dimensional information feature results of the target are input into the trained convolutional neural network model to obtain the target classification results and complete the signal processing for multi-dimensional radar target intelligent recognition.

[0046] This invention primarily utilizes deep learning networks to process target feature results. The approach to processing radar data using deep learning networks mainly involves in-depth analysis of the data by professionals to extract features, thereby establishing a recognition feature library; constructing a multi-level machine learning model, and using automatic layer-by-layer feature transformation by the computer to extract the inherent features of the input data; and automatically extracting features from the feature model using massive training data, thereby forming classification constraints based on the training data information, ultimately improving the accuracy of classification or prediction.

[0047] Completing the above steps constitutes the signal processing method for intelligent recognition of multi-dimensional radar targets.

[0048] Those skilled in the art will recognize that the described embodiments are intended to help readers understand the principles of the invention and should be understood as not limiting the scope of protection of the invention to the described embodiments. Various modifications and variations can be made to the invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the scope of the claims of the invention.

Claims

1. A signal processing method for multi-dimensional radar target intelligent recognition, characterized in that, Includes the following steps: Step 1, radar transmit waveform modulation. The radar transmit waveform uses multiple modulation methods to achieve the transmission of multi-modulated signals, including narrowband modulation, wideband modulation, and long-time frame accumulation modulation. Narrowband modulation includes frequency modulation, phase modulation, and pulse train modulation; wideband modulation includes frequency modulation, phase modulation, pulse train modulation, and composite modulation; long-time frame-accumulation modulation includes step frequency modulation and space-time-frequency coded composite modulation. Step 2, demodulation of received signal: The radar demodulates the received signal according to the modulation method of the transmitted signal. The echo signal generated when the radar transmits the modulated signal and encounters the target is received by the antenna, amplified by radio frequency, frequency converted, demodulated and analog-to-digital converted to form the received signal. The demodulation transformation is performed on the received signal after long-term accumulation, code correlation, frequency modulation demodulation, and phase demodulation to form a signal processing and acquisition digital signal; all targets are known targets, and the target category is recorded as the truth value label of the target; Step 3, multi-mode signal processing, through spectrum analysis, time-frequency analysis, range correlation demodulation, long-term accumulation spectrum analysis, high-resolution range demodulation, multi-dimensional spectrum analysis, and time-range-Doppler transform processing, to obtain the multi-dimensional information feature results of the target; Step 4: Deep learning target feature extraction. Deep learning network training is carried out using convolutional neural networks to target multidimensional information feature results. The network layer parameters are associated with the target features, and the trained convolutional neural network model is saved. Step 5, multi-mode recognition: Through parallel / time-division multi-mode signal processing, the transmitted waveform of the radar signal is modulated, the received signal is demodulated, and the multi-dimensional information feature results of the target are obtained. The multi-dimensional information feature results of the target are input into the trained convolutional neural network model to obtain the target classification result, thus completing the signal processing for multi-dimensional radar target intelligent recognition.

2. The signal processing method for multi-dimensional radar target intelligent recognition according to claim 1, characterized in that, The multi-modulation signals in step 1 include linear frequency modulated radar signals, nonlinear frequency modulated radar signals, polynomial phase modulated radar signals, sinusoidal frequency modulated radar signals, phase shift keying radar signals, binary phase coded and quadrature phase coded radar signals, frequency coded radar signals, linear frequency modulated-phase coded radar signals, frequency coded-phase coded radar signals, step frequency modulated signals, and space-time frequency coded signals.

3. The signal processing method for multi-dimensional radar target intelligent recognition according to claim 1, characterized in that, The specific method for step 3 is as follows: Step 301: Spectrum analysis. Using conventional FFT transformation, the velocity, energy, and micro-doppler information of the target are obtained under narrowband modulation signal. Step 302: Time-frequency analysis. Short-time FFT, WVD, and wavelet transform are used to obtain the time-frequency characteristic information of the target under narrowband modulation signal. Step 303: Range-related demodulation, including frequency modulation demodulation and phase modulation demodulation, to obtain the target's range information parameters under narrowband modulation signal; combined with the energy information characteristics of the spectrum analysis results, to calculate the corresponding range echo energy and characterize the target's radar cross-section. Step 304: Long-term accumulated spectrum analysis mainly relies on multi-frame continuous processing. It processes the long-term accumulated frame modulation signal to obtain the time-frequency transformation characteristics, micro-doppler, and motion trend characteristics of the target over a long period of time. Step 305: High-resolution range demodulation. By using the range demodulation results of down-modulation and phase-modulation demodulation of the broadband modulated signal, high-resolution one-dimensional range image information and high-precision range change feature information of the target are obtained. Step 306: Multidimensional spectrum analysis. For long-term accumulated processing, multidimensional spectrum analysis calculation is performed to obtain the target micro-motion Doppler information and target motion trend characteristic information. Step 307: Time-range-Doppler transformation, for long-term accumulation processing and signal composite processing, to obtain the target's time, energy, distance, and Doppler multidimensional composite information.

4. The signal processing method for multi-dimensional radar target intelligent recognition according to claim 1, characterized in that, The specific method for step 4 is as follows: Step 401: Input data, with a dimension of batch_size × input_dim; batch_size is the amount of data used in the training iteration, and input_dim is the dimension of the input layer data, that is, the dimension of the multidimensional information feature result of the target. Step 402: Construct a convolutional neural network architecture, including an input layer, a hidden layer, and an output layer. The input layer has a data dimension of input_dim and an output dimension of output_dim. The hidden layer has an input dimension of output_dim and an output dimension of num_feature. The output layer has an input dimension of num_feature and an output dimension of num_classes. Step 403, forward propagation, inputting the input data into the convolutional neural network, respectively passing through the input layer, the hidden layer and the output layer, and outputting a target classification result ; Step 404, calculate loss, compare the target classification result of the forward output The true value label corresponding to the input data The loss function value between the two is calculated by using the cross-entropy loss function ; Step 405, back propagation, according to the calculated loss function value on the parameters in the convolutional neural network optimization; Step 406: Repeat steps 403-405. When the loss function value no longer decreases for several consecutive rounds, stop the loop and obtain the final convolutional neural network model.

Images