A MIMO radar single target identification method based on CV-BP network

Abstract

The application discloses a MIMO radar single-target identification method based on a CV-BP network and a neural network training method. The CV-BP neural network with complex values of activation functions, weights and biases is constructed through MIMO radar to collect target complex echo data, the complex echo data is directly input into the network for training to obtain a target identification model, and the classification and identification of the target are realized. The amplitude and phase information of the complex data are completely retained, and the accuracy and robustness of target identification are improved.

Description

Technical Field

[0001] This invention belongs to the field of radar target recognition technology, specifically relating to a MIMO radar single target recognition method and neural network training method based on CV-BP network. Background Technology

[0002] MIMO (Multiple-Input Multiple-Output) radar systems, by configuring multiple transmitting and receiving antennas, can simultaneously detect targets from different angles, thereby acquiring richer target information. Compared with traditional single-input single-output radars, MIMO radars have higher spatial resolution, better parameter identification, and stronger anti-jamming capabilities, and have broad application prospects in military reconnaissance, civilian monitoring, autonomous driving, and other fields.

[0003] In radar target identification tasks, target echo data is typically in complex-valued form, containing both amplitude and phase information. Traditional target identification methods usually require converting the complex-valued echo data into real-valued data before processing, such as extracting amplitude information or processing the real and imaginary parts separately. This processing method inevitably leads to the loss of phase information or the destruction of the integrity of the complex-valued data structure, reducing the accuracy and reliability of target identification.

[0004] Complex echo data from airborne radar contains rich target features. The amplitude and phase of the complex data represent energy magnitude and temporal or spatial correspondence, respectively, exhibiting good robustness. However, most existing radar target recognition methods borrow from optical methods, using real-valued neural networks to classify and recognize radar images. For complex echo data processing, they mostly extract the magnitude of the complex data, utilizing only the amplitude information and ignoring the phase information. Furthermore, even when using complex data, the real and imaginary parts are often input as two independent features into the network for training, without considering the relationship between the real and imaginary parts. Most radar target recognition methods require human intervention to extract and select target features. The recognition effect largely depends on the results of feature extraction and selection. Currently, there is no unified framework for feature extraction methods and the extracted features; they are mostly determined based on the specific target category and the form of the target echo. This leads to different recognition results when different features of the same target are selected for training, exhibiting subjectivity and significant uncertainty. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a MIMO radar single target recognition method and neural network training method based on CV-BP network. This method can directly use the complex-valued echo data collected by MIMO radar as network input, fully preserve amplitude and phase information, and improve the accuracy and efficiency of target recognition.

[0006] To achieve the above objectives, a first aspect of the present invention provides a MIMO radar single-target identification method based on a CV-BP network, comprising: The target scene is scanned using a MIMO radar to collect complex-valued echo data of the target. The complex-valued echo data contains target amplitude and phase information obtained from multiple transmission and reception paths of the MIMO radar. Construct a CV-BP neural network, which includes an input layer, at least one hidden layer, and an output layer. The activation function, weights, and bias parameters of the CV-BP neural network are all complex values. The complex-valued echo data is directly used as the input of the CV-BP neural network. The network is trained through forward propagation and error backpropagation to obtain the trained target recognition model. The trained target recognition model is used to classify and identify the complex-valued echo data of the target to be identified.

[0007] Furthermore, the number of sampling points for the transmitted waveform of the MIMO radar is P, the number of azimuth scanning points is Q, and the data set size corresponding to a single target is P×Q×N×M, where N and M are the number of transmitting antennas and receiving antennas of the MIMO radar, respectively.

[0008] Furthermore, before constructing the target recognition model, the method further includes: The complex-valued echo data of different types of targets are stacked to obtain the total dataset. The echo data of a single target are shuffled and divided into a training set, a validation set, and a test set. The test set accounts for 20% of the total echo data, the validation set accounts for 20% of the training data, and the remaining data is used as the training set.

[0009] Furthermore, the CV-BP neural network includes an input layer, two hidden layers, and an output layer. The hidden layers use the complex ReLU activation function to activate the real and imaginary parts of the input vector respectively through an element-wise operation algorithm. The output layer uses the complex Softmax activation function.

[0010] Furthermore, the CV-BP neural network uses the complex cross-entropy loss function to measure the error between the complex output signal and the target label.

[0011] Furthermore, the partial derivatives of the complex cross-entropy loss function with respect to the complex weights and complex biases are calculated according to the chain rule, and the complex weights and complex biases are iteratively updated based on the partial derivatives. Training stops when the output error reaches a preset threshold or the number of iterations reaches a set value.

[0012] A second aspect of the present invention provides a CV-BP neural network training method for MIMO radar complex-valued echo data, comprising: Acquire target complex-valued echo data collected by MIMO radar, wherein the complex-valued echo data contains target amplitude and phase information acquired by multiple transmission and reception paths of the MIMO radar; The complex-valued echo data is directly input into the CV-BP neural network for forward propagation to obtain a complex-valued output signal. The activation function, weights, and bias parameters of the CV-BP neural network are all complex values. The real and imaginary parts of each input vector are activated sequentially through an element-wise algorithm. The error between the complex-valued output signal and the target label is calculated using the complex cross-entropy loss function. The partial derivatives of the complex cross-entropy loss function with respect to the complex weights and complex biases are calculated according to the chain rule. Based on the partial derivatives, the complex weights and complex biases are iteratively updated according to the update formulas for the weights and biases until the training stopping condition is met.

[0013] Furthermore, in the aforementioned forward propagation, the first l The complex output signal of the i-th neuron in layer +1 is calculated by the complex weights, complex biases and complex activation functions of that layer. The complex activation function of the hidden layer is the complex ReLU activation function and the complex activation function of the output layer is the complex Softmax activation function.

[0014] Furthermore, the update formula incorporates a learning rate to control the parameter update speed, and the training stopping condition is that the output error between the output value and the target value reaches a preset threshold or the number of iterations reaches a set value.

[0015] Furthermore, the CV-BP neural network is a four-layer network structure, including one input layer, two hidden layers and one output layer, with corresponding complex neuron numbers of 101, 100, 32 and 4, respectively. The batch size is 600, the learning rate is 0.01, and the optimizer uses stochastic gradient descent.

[0016] This application provides a MIMO radar single-target recognition method and neural network training method based on a CV-BP network. The recognition method includes: scanning a target scene using a MIMO radar to collect complex-valued echo data of the target, wherein the complex-valued echo data contains target amplitude and phase information obtained from multiple transmission and reception paths of the MIMO radar; constructing a CV-BP neural network, wherein the CV-BP neural network includes an input layer, at least one hidden layer, and an output layer, and the activation function, weights, and bias parameters of the CV-BP neural network are all complex values; directly using the complex-valued echo data as the input of the CV-BP neural network, and training the network through forward propagation and error backpropagation to obtain a trained target recognition model; and using the trained target recognition model to classify and recognize the complex-valued echo data of the target to be identified. This application has the following technical effects: (1) The present invention directly uses the complex-valued echo data collected by MIMO radar as the input of CV-BP neural network without converting the complex-valued data into real-valued data, thus preserving the amplitude and phase information in the echo signal, avoiding information loss in the data preprocessing process, and improving the accuracy of target recognition.

[0017] (2) In the CV-BP neural network constructed in this invention, the activation function, weights and bias parameters are all complex values, which can directly perform feature extraction and classification operations in the complex domain, which is more in line with the inherent structural characteristics of MIMO radar complex-valued echo data, and enhances the network's ability to express complex-valued signals and its learning efficiency. Attached Figure Description

[0018] Figure 1 The actual test scenario provided for the embodiments of the present invention includes a turntable, a radar platform, and a computer; Figure 2 The measured object provided in the embodiment of the present invention, wherein, Figure 2 (a) is a kettle. Figure 2 (b) is a toy car. Figure 2 (c) is a protrusion. Figure 2 (d) is the computer case; Figure 3 A multilayer CV-BP network structure is provided in an embodiment of the present invention; Figure 4 A flowchart of a MIMO radar single-target recognition method based on a CV-BP network provided in an embodiment of the present invention. Detailed Implementation

[0019] At the very beginning of the detailed embodiments, the key terms involved in this invention are explained: CV-BP: Complex-Valued Back-Propagation; MIMO: Multiple-input multiple-output. ReLU: Rectified Linear Unit; SGD: Stochastic Gradient Descent.

[0020] Example 1 As attached Figure 1 As shown, attached Figure 1 This is a real-world testing scenario, including a turntable, radar platform, and computer. The MIMO radar scans the target scene, acquiring complex-valued echo data of the target. This data contains target amplitude and phase information obtained from multiple transmission and reception paths of the MIMO radar. Its purpose is to utilize the multiple observation channels of the MIMO radar, which coexist in space, to provide a large amount of spatiotemporal information containing different amplitudes, delays, or phases of the target, thus acquiring more target information.

[0021] As attached Figure 4 As shown, Figure 4 This is a flowchart of a MIMO radar single-target recognition method based on a CV-BP network provided in an embodiment of the present invention. The method may include: Step S101: Use MIMO radar to scan the target scene and collect the complex-valued echo data of the target. The complex-valued echo data contains the target amplitude and phase information obtained by multiple transmission and reception paths of MIMO radar. In one possible implementation, the number of sampling points for the transmitted waveform of the MIMO radar is P, the number of azimuth scanning points is Q, and the dataset size corresponding to a single target is P,QNM, where N and M are the number of transmitting antennas and receiving antennas of the MIMO radar, respectively.

[0022] In the embodiments provided by this invention, the data set dimension is determined by setting a specific number of sampling points and azimuth scanning points, combined with the number of antennas. Here, P corresponds to the number of sampling frequency points.

[0023] In one possible implementation, before constructing the target recognition model, the method further includes: stacking the complex-valued echo data of different types of targets to obtain a total dataset, and shuffling the echo data of a single target body to divide it into a training set, a validation set, and a test set, wherein the test set accounts for 20% of the total echo data, the validation set accounts for 20% of the training data, and the remaining data is used as the training set.

[0024] In the embodiments provided by this invention, the extracted radar echo data of different types are stacked to obtain the total dataset of MIMO radar, and the dataset is then divided for training deep learning models. Specifically, the data of individual targets needs to be shuffled before being randomly allocated.

[0025] For example, the AWR1642 radar chip (containing 2 transmit channels or 4 receive channels) is used to collect target echo data, such as... Figure 1 As shown. The specific measurement process is as follows: First, connect the radar platform to the computer and power supply respectively. Then, place the target in the center of the turntable (try to ensure that the center of the target is located in the center of the turntable), and align it with the reference point of the radar platform. Finally, configure the parameters of the radar transmission signal on the computer. After the settings are completed, data acquisition can begin. During the acquisition process, the radar acquires data once after each rotation of the turntable, i.e., "one stop, one measurement". The rotation angle range is... Sampling interval is taken During the actual data acquisition, the distance from the antenna on the AWR1642 board to the desktop was 0.10m, and the distance to the center of the turntable was approximately 0.84m. Repeating the above measurement process, echo data were finally collected from four types of targets: a kettle (base with a diameter of 0.16m and a height of 0.20m), a toy car (length 0.26m, width 0.09m, and height 0.15m), a protrusion (base with a diameter of 0.16m and a height of 0.20m), and a computer case (length 0.42m, width 0.41m, and height 0.18m). See the attached image for the actual objects. Figure 2 .

[0026] Stacking the four types of radar echo data extracted above yields the total dataset [200, 601] for the MIMO radar. 8 4] = [200, 19232]. The created dataset is divided into training / validation sets and test sets for training the deep learning model. The data allocation is as follows: the test set accounts for 20% of the total echo data, the validation set accounts for 20% of the total training data during training, and the remaining data is the training set. Here, the data for each single target needs to be shuffled before being randomly allocated.

[0027] Step S102: Construct a CV-BP neural network. The CV-BP neural network includes an input layer, at least one hidden layer, and an output layer. The activation function, weights, and bias parameters of the CV-BP neural network are all complex values. As attached Figure 3 As shown, attached Figure 3It is a multi-layer CV-BP network structure, consisting of an input layer, multiple hidden layers, and an output layer. The CV-BP neural network is constructed by adding an input layer, at least one hidden layer, and an output layer. The activation function, weights, and bias parameters of the CV-BP neural network are all complex values. Its purpose is to extend the BP neural network from the real-number domain to the complex-number domain, constructing a network model capable of directly processing complex-valued data.

[0028] In one possible implementation, the CV-BP neural network includes an input layer, two hidden layers, and an output layer. The hidden layers use the complex ReLU activation function to activate the real and imaginary parts of the input vector respectively through an element-wise operation algorithm, and the output layer uses the complex Softmax activation function.

[0029] In the embodiments provided by this invention, the hidden layers are activated using the complex ReLU function, and the output layer is activated using the complex Softmax function. Specifically, the real and imaginary parts of each input vector are activated sequentially using an element-wise algorithm.

[0030] In one possible implementation, the CV-BP neural network uses a complex cross-entropy loss function to measure the error between the complex-valued output signal and the target label.

[0031] In the embodiments provided by the present invention, the activation function of the network output layer is the complex Softmax function, and the error function is measured by the complex cross-entropy loss function.

[0032] In one possible implementation, the partial derivatives of the complex cross-entropy loss function with respect to the complex weights and complex biases are calculated according to the chain rule, and the complex weights and complex biases are iteratively updated based on the partial derivatives. Training stops when the output error reaches a preset threshold or the number of iterations reaches a set value.

[0033] In the embodiments provided by this invention, the parameters of the CV-BP algorithm are updated by minimizing the loss function of the training data, and the update process of the complex weight values ​​and complex bias values ​​is given according to the chain rule. Specifically, network training stops when the output error between the output value and the target value reaches a certain threshold or when the set number of iterations is reached.

[0034] For example, suppose the input layer of the network has P neurons, and the input vector can be represented as: , Let P be a complex number, corresponding to the number of sampling frequency points, with superscripts R and I representing the real and imaginary parts, respectively. In the (l+1)th layer, the complex output signal of the i-th neuron can be expressed as... (1) in, This is the complex output of the i-th neuron in the (l+1)-th layer. This is the complex output signal of the output layer. and These represent the ReLU function and the Softmax activation function, respectively. Here, the real and imaginary parts of each input vector are activated sequentially using an element-wise algorithm.

[0035] The total output value of the output layer can be expressed as In training a model, the network's output value is often difficult to achieve the expected result in a single training iteration. Therefore, it is necessary to iterate multiple times to gradually modify the output value, minimizing the error between the output value and the actual target value. Here, the desired complex target value (i.e., the true value to be fitted) is defined as... The desired output signal can then be expressed as .

[0036] In multi-class classification problems, one-hot encoding is often used to assign a label to each target category, with the label length equal to the number of target categories. The activation function of the network output layer is the complex softmax function, and the corresponding error function is the complex cross-entropy (CE) loss function. Assuming there is only one sample, then... (2) Where N is the number of samples in the training set.

[0037] The parameters of the CV-BP algorithm can be updated by minimizing the loss function of the training data. Here, based on the chain rule, the update process for the complex weights and complex biases is given. Error function. about The partial derivatives are (3) Similarly, error function about The partial derivatives are (4) in, , .

[0038] The update formulas for weights and biases are as follows: (5) In the formula, The learning rate controls how quickly the parameters are updated. Network training stops when the output error between the output value and the target value reaches a certain threshold or when the set number of iterations is reached.

[0039] Step S103: The complex-valued echo data is directly used as the input of the CV-BP neural network. The network is trained through forward propagation and backpropagation of error to obtain the trained target recognition model. Complex-valued echo data is directly used as input to a CV-BP neural network. The network is trained through forward propagation and error backpropagation to obtain a trained target recognition model. Its purpose is to directly learn features using complex-valued data containing amplitude and phase information, avoiding the subjectivity and uncertainty associated with manual feature extraction.

[0040] Step S104: Use the trained target recognition model to classify and recognize the complex-valued echo data of the target to be identified.

[0041] The trained target recognition model is used to classify and identify the complex-valued echo data of the target to be identified. Its purpose is to achieve accurate identification of targets in the surrounding environment under various lighting and weather conditions.

[0042] This embodiment provides a single-target recognition method for MIMO radar based on a CV-BP network. It extends the BP neural network from the real-number domain to the complex-number domain, proposes a CV-BP learning algorithm, and applies it to the single-target classification and recognition problem of MIMO radar. The CV-BP network can autonomously learn the characteristics of the target, fully utilizing the amplitude and phase information in the complex echo data, effectively improving the target recognition rate, and exhibiting good robustness to noise. The multi-observation channel characteristic of MIMO radar allows it to obtain more echo data of different amplitudes and phases of the target, resulting in a better trained network model and better performance at low signal-to-noise ratios.

[0043] Example 2 In one possible implementation, during forward propagation, the first... l The complex output signal of the i-th neuron in layer +1 is calculated using the complex weights, complex biases, and complex activation functions of that layer. The complex activation function of the hidden layer is the complex ReLU activation function, and the complex activation function of the output layer is the complex Softmax activation function.

[0044] In the embodiments provided by the present invention, the derivation of the forward propagation algorithm is carried out layer by layer. Complex value data enters the network from the input layer, and then is processed by each hidden layer before finally reaching the output layer.

[0045] The complex cross-entropy loss function is used to calculate the error between the complex-valued output signal and the target label. Its purpose is to accurately measure the error between the complex-valued network output and the true value in multi-class classification problems.

[0046] The partial derivatives of the complex cross-entropy loss function with respect to the complex weights and biases are calculated using the chain rule. Based on these partial derivatives, the complex weights and biases are iteratively updated according to the update formulas for the weights and biases until the training stopping condition is met. Its function is to continuously adjust and optimize the model parameters through error backpropagation.

[0047] In one possible implementation, a learning rate is introduced into the update formula to control the speed of parameter updates, and the training stops when the output error between the output value and the target value reaches a preset threshold or the number of iterations reaches a set value.

[0048] In the embodiments provided by the present invention, a learning rate is introduced into the update formulas for weights and biases to control the update speed of the parameters.

[0049] In one possible implementation, the CV-BP neural network is a four-layer network structure, including one input layer, two hidden layers and one output layer, with the corresponding number of complex neurons being 101, 100, 32 and 4, respectively. The batch size is 600, the learning rate is 0.01, and the optimizer is the stochastic gradient descent method.

[0050] For example, in radar target recognition, the recognition rate (also known as accuracy, which refers to the proportion of samples that are correctly identified as the sample category out of the total number of samples) is often used as a network evaluation indicator.

[0051] Here, the network structure and activation function of R-BP and CV-BP remain unchanged, only the number of neurons in the input layer is changed to 200 and 400 respectively. This is because the R-BP network only uses amplitude or phase frequency data as input. Table 1 shows the target recognition results of CV-BP and R-BP algorithms under MIMO radar measured data.

[0052] Table 1 Target recognition results from MIMO radar measured data

[0053] As shown in Table 5-2, for the real-valued BP network, the recognition rates using the amplitude and phase frequency responses of the echo data as input features are 99.76% and 99.06%, respectively, effectively identifying the four types of targets. For the CV-BP network, directly inputting complex-valued data into the network as features results in a target recognition rate of 99.96%, also accurately distinguishing the four types of targets.

[0054] To verify the robustness of the CV-BP network to noise, Gaussian white noise was added to the echo data (signal-to-noise ratio intensity varied from -10dB to 20dB, with a step interval of 5dB), while other conditions remained unchanged. Table 2 shows the target recognition results obtained by adding noise of different intensities to the measured MIMO radar data, where C, A, and P correspond to complex echo data, amplitude frequency data, and phase frequency data, respectively.

[0055] Table 2. Target recognition results after adding noise to MIMO radar measured data.

[0056] Table 2 shows that for both the R-BP network using amplitude and phase frequency data and the CV-BP network using complex-valued echo data, the target classification and recognition performance increases with the increase of signal-to-noise ratio. When 5 dB of Gaussian white noise is added, the recognition rates of R-BP(A) and R-BP(P) decrease from 99.76% and 99.06% under ideal conditions to 86.85% and 94.15%, respectively, while the recognition rate of CV-BP(C) only changes from 99.96% to 99.35%.

Claims

1. A MIMO radar single-target identification method based on CV-BP network, characterized in that, include: The target scene is scanned using a MIMO radar to collect complex-valued echo data of the target. The complex-valued echo data contains target amplitude and phase information obtained from multiple transmission and reception paths of the MIMO radar. Construct a CV-BP neural network, which includes an input layer, at least one hidden layer, and an output layer. The activation function, weights, and bias parameters of the CV-BP neural network are all complex values. The complex-valued echo data is directly used as the input of the CV-BP neural network. The network is trained through forward propagation and error backpropagation to obtain the trained target recognition model. The trained target recognition model is used to classify and identify the complex-valued echo data of the target to be identified.

2. The MIMO radar single-target identification method based on CV-BP network according to claim 1, characterized in that, The MIMO radar has P sampling points for the transmitted waveform, Q azimuth scanning points, and a dataset size of P,QNM for a single target, where N and M are the number of transmitting antennas and receiving antennas of the MIMO radar, respectively.

3. The MIMO radar single-target identification method based on CV-BP network according to claim 1, characterized in that, Before constructing the target recognition model, the method further includes: The complex-valued echo data of different types of targets are stacked to obtain the total dataset. The echo data of a single target are shuffled and divided into a training set, a validation set, and a test set. The test set accounts for 20% of the total echo data, the validation set accounts for 20% of the training data, and the remaining data is used as the training set.

4. The MIMO radar single-target identification method based on CV-BP network according to claim 1, characterized in that, The CV-BP neural network includes an input layer, two hidden layers, and an output layer. The hidden layers use the complex ReLU activation function to activate the real and imaginary parts of the input vector respectively through an element-wise operation algorithm. The output layer uses the complex Softmax activation function.

5. The MIMO radar single-target identification method based on CV-BP network according to claim 1, characterized in that, The CV-BP neural network uses the complex cross-entropy loss function to measure the error between the complex output signal and the target label.

6. The MIMO radar single-target identification method based on CV-BP network according to claim 5, characterized in that, The partial derivatives of the complex cross-entropy loss function with respect to the complex weights and complex biases are calculated according to the chain rule, and the complex weights and complex biases are iteratively updated based on the partial derivatives. Training stops when the output error reaches a preset threshold or the number of iterations reaches a set value.

7. A CV-BP neural network training method for MIMO radar complex-valued echo data, characterized in that, include: Acquire target complex-valued echo data collected by MIMO radar, wherein the complex-valued echo data contains target amplitude and phase information acquired by multiple transmission and reception paths of the MIMO radar; The complex-valued echo data is directly input into the CV-BP neural network for forward propagation to obtain a complex-valued output signal. The activation function, weights, and bias parameters of the CV-BP neural network are all complex values. The real and imaginary parts of each input vector are activated sequentially through an element-wise algorithm. The error between the complex-valued output signal and the target label is calculated using the complex cross-entropy loss function. The partial derivatives of the complex cross-entropy loss function with respect to the complex weights and complex biases are calculated according to the chain rule. Based on the partial derivatives, the complex weights and complex biases are iteratively updated according to the update formulas for the weights and biases until the training stopping condition is met.

8. The CV-BP neural network training method for MIMO radar complex-valued echo data according to claim 7, characterized in that, In the forward propagation, the first l The complex output signal of the i-th neuron in layer +1 is calculated by the complex weights, complex biases and complex activation functions of that layer. The complex activation function of the hidden layer is the complex ReLU activation function and the complex activation function of the output layer is the complex Softmax activation function.

9. The CV-BP neural network training method for MIMO radar complex-valued echo data according to claim 7, characterized in that, The update formula incorporates a learning rate to control the parameter update speed, and the training stopping condition is that the output error between the output value and the target value reaches a preset threshold or the number of iterations reaches a set value.

10. The CV-BP neural network training method for MIMO radar complex-valued echo data according to claim 7, characterized in that, The CV-BP neural network is a four-layer network structure, including one input layer, two hidden layers and one output layer, with corresponding complex neuron numbers of 101, 100, 32 and 4, respectively. The batch size is 600, the learning rate is 0.01, and the optimizer uses stochastic gradient descent.

Images