Radar target HRRP identification method based on quantum hybrid neural network

Abstract

The invention belongs to the technical field of radar target recognition, and particularly relates to a radar target HRRP recognition method based on a quantum hybrid neural network. According to the method, firstly, a three-layer convolutional neural network (CNN) is utilized to extract multi-level local features of a one-dimensional image, then the multi-level local features are input into a quantum circuit model to further extract entanglement features as classification features, and finally the category of a target is determined. The quantum entanglement characteristics reflect the related information between adjacent data elements, and the entanglement characteristics in the noise data are very weak, so that the influence of noise on the recognition performance is reduced, and the recognition rate under the condition of low signal-to-noise ratio is improved. Recognition experiments are carried out by using the HRRP data sets of the two types of aircraft targets, and results show that the recognition rate of the method reaches 89% under the condition that the signal-to-noise ratio is 0dB.

Description

Technical Field

[0001] The invention belongs to the technical field of radar target recognition, and in particular relates to a radar target HRRP recognition method based on quantum hybrid neural network. Background Art

[0002] The radar one-dimensional profile (HRRP) is an amplitude spectrum formed by coherent accumulation of complex echoes of target scatterers in each range unit. HRRP can reflect the geometric shape and internal structure characteristics of the target in the direction of the radar line of sight, and it is easy to obtain and process. Therefore, the HRRP recognition method plays an important role in radar target recognition.

[0003] In target recognition methods, effective feature extraction is the key. For example, recognition methods based on HRRP fast Fourier transform amplitude features, high-order spectrum features, feature subspaces, and principal component analysis. These methods have achieved good recognition results, but the recognition features need to be manually determined based on experience. Deep learning methods can automatically learn multi-level features from data, thereby improving recognition results. Studies have shown that deep learning methods such as convolutional neural networks (CNNs), long short-term memory networks (LSTMs), residual networks, and recurrent neural networks (RNNs) have significantly better recognition rates than the above traditional methods. However, these deep learning methods are very sensitive to noise, resulting in a significant decrease in recognition performance under low signal-to-noise ratio conditions. Summary of the invention

[0004] In view of the above problems, the present invention proposes a quantum hybrid neural network method, which can extract the entanglement characteristics in the data, thereby improving the recognition performance.

[0005] The technical solution of the present invention is:

[0006] 1. A radar target HRRP recognition method based on quantum hybrid neural network, comprising the following steps:

[0007] S1, obtain multiple HRRP sample data of multiple categories of targets to build a training data set;

[0008] S2. Construct a quantum hybrid neural network for target recognition, including a CNN subnet and a quantum subnet.

[0009] The CNN subnet is used to extract features from the input HRRP sample data to obtain convolution features matching the quantum subnet;

[0010] The quantum subnet includes a quantum circuit and a classification layer, wherein the quantum circuit includes a first quantum circuit and a second quantum circuit, and the two quantum circuits each include an H gate, a rotating Y gate, a rotating Z gate and a measurement gate connected in sequence, and a CNOT gate shared by the two quantum circuits is provided between the rotating Z gate and the measurement gate; wherein the rotation angles θ of the rotating Y gate and the rotating Z gate are obtained by convolution features;

[0011] The H gate is used to convert the input quantum bit from the ground state to a superposition state with equal probability, expressed as:

[0012]

[0013] Among them, |Q 01 >and|Q 11 > respectively represent the quantum superposition state of the first quantum circuit and the quantum superposition state of the second quantum circuit, |0> and |1> respectively represent the two ground states of quantum;

[0014] The rotating Y gate is used for the quantum superposition state |Q 01 >and|Q 11 >Perform the operation:

[0015]

[0016] i=0,1

[0017] Among them, θ i represents the rotation angle of the rotating Y gate, i=0 means the parameter corresponds to the first quantum circuit, i=1 means the parameter corresponds to the second quantum circuit;

[0018] The rotating Z gate pairs |Q i2 >Processing yields:

[0019]

[0020]

[0021] Among them, |Q 03 > represents the rotational Z gate output of the first quantum circuit, θ2 is the rotation angle of the rotational Z gate of the first quantum circuit, |Q 13 > represents the rotational Z gate output of the second quantum circuit, θ3 is the rotation angle of the rotational Z gate of the second quantum circuit;

[0022] The input of the CNOT gate is |Q 03 >and|Q 13 >, the output is:

[0023]

[0024]

[0025]

[0026]

[0027]

[0028] The measurement gate performs a quantum measurement on |Q4> and obtains:

[0029] P ij =<Q4|M ij |Q4>

[0030] where M ij is the measurement operator, M ij =|ij><ij|, |ij> represents the computational basis state, <ij| is the conjugate transpose of |ij>, i,j = 0,1;

[0031] After the quantum measurement, the probability values of |Q4> under the four computational bases |00>, |01>, |11>, |10> are obtained, and then input into the classification layer for classification;

[0032] S3. Use the training data set obtained in S1 to train the quantum hybrid neural network constructed in S2. The training process is optimized by using the least mean square error loss function and the steepest gradient descent optimization method to obtain the trained quantum hybrid neural network;

[0033] S4. Input the HRRP sample to be recognized into the trained quantum hybrid neural network, and use the label corresponding to the largest component in the output vector of the classification layer as the target recognition category.

[0034] Furthermore, the CNN subnet is composed of three convolutional layers, a pooling layer, and a fully connected layer connected in sequence. Among them, the convolutional layer consists of a convolutional operation, batch normalization, and ReLu operation. The sizes of the convolutional kernels in the three convolutional layers are all 1*3, and the numbers of channels are 32, 64, and 64 respectively.

[0035] The beneficial effect of the present invention is that since the quantum entanglement feature reflects the correlation information between adjacent data elements, and the entanglement characteristic in the noise data is very weak, the method of the present invention reduces the influence of noise on the recognition performance, thereby improving the recognition rate under low signal-to-noise ratio conditions. Recognition experiments are carried out using the HRRP data sets of two types of aircraft targets. The results show that when the signal-to-noise ratio is 0dB, the recognition rate of the method of the present invention reaches 89%. Description of the Drawings Figure 1 is the structural diagram of the quantum hybrid neural network of the present invention; Figure 2 is the structural diagram of the convolutional subnet; Figure 3It is the structural diagram of the convolution layer; Figure 4 This is a diagram of the quantum subnet structure. DETAILED DESCRIPTION

[0036] The present invention will be further described below in conjunction with the accompanying drawings.

[0037] The structure of the quantum hybrid neural network proposed in the present invention is as follows Figure 1 As shown, it consists of a convolutional subnet and a quantum subnet.

[0038] Due to the limited number of quantum bits, the CNN convolutional subnet first performs feature extraction and dimensionality reduction, and then the quantum subnet performs entanglement feature extraction and classification. The structures of the CNN convolutional subnet and the quantum subnet are as follows.

[0039] (1) Structure of CNN subnet

[0040] The structure of the convolutional subnet is as follows Figure 2 As shown in Figure 1, it consists of three convolutional layers, a pooling layer and a fully connected layer, where the structure of the convolutional layer is as follows: Figure 3 As shown in the figure, it consists of convolution operation, batch normalization and ReLu operation. The size of the convolution kernel in the three convolution layers is 1*3, and the number of channels is 32, 64 and 64 respectively. The convolution layer uses convolution operation to generate different convolution features, while batch normalization can accelerate network convergence, and then the nonlinearity is introduced through ReLu operation to further enrich the features. Then the convolution features are input to the pooling layer for maximum sampling to achieve feature compression. Finally, the fully connected layer is used to output the features to match the input requirements of the quantum subnet.

[0041] (2) Quantum subnet structure

[0042] The quantum subnet structure is as follows Figure 4 As shown in Figure 1, the quantum subnet consists of a quantum circuit and a classification layer. The quantum circuit mainly relies on quantum gates for operation, including H gates, rotating Y gates, rotating Z gates, and CNOT gates. 00 >and|Q 10 > represents the initial quantum bit, |Q i1 >、|Q i2 >、|Q i3 >(i=0,1) respectively indicates that the i-th quantum bit passes through the H gate, R gate, Y (θ) gate (rotating Y gate), R Z (θ) The quantum state after the gate (rotating Z gate). At the same time θ i (i=0,1,2,3) represents the convolutional features output by the CNN subnet. The classification layer consists of a fully connected layer and a SoftMax classifier.

[0043] First, the qubit is initialized through the H gate. The H gate initializes |Q 00>|Q 10 >is converted from the ground state |0> to a superposition state with equal probabilities:

[0044]

[0045] where |0> and |1> represent the two ground states of the quantum respectively. Then, taking θ0 and θ1 as the rotation angles of the R Y (θ) gate, operating on the quantum superposition state |Q 01 >|Q 11 >results in:

[0046]

[0047] Next, the R Z (θ) gate is introduced, with rotation angles θ2 and θ3. Then the expression of |Q 03 >is as follows:

[0048]

[0049] Similarly, the expression of |Q 13 >is as follows:

[0050]

[0051] Then, through the CNOT gate, the quantum entangled state |Q4> is formed:

[0052]

[0053] where

[0054]

[0055]

[0056]

[0057]

[0058] Finally, quantum measurement is performed. Let the measurement operator be M ij = |ij><ij| (i, j = 0, 1), where |ij> represents the computational ground state and <ij| is the conjugate transpose of |ij>. Then the probability P ij of |Q4> under |ij> is:

[0059] P ij = <Q4|M ij |Q4> (10)

[0060] That is, the conjugate transpose <Q4| of |Q4>, M ijThe product of |Q4> and |Q4>. After quantum measurement, the probability value of |Q4> in four calculation bases |00>, |01>, |11>, |10> is obtained. Then it is input into the classification layer for classification.

[0061] The training sample set of the one-dimensional image of the radar target is used as input. The entanglement features are extracted by the CNN subnet and the quantum subnet. The corresponding category is determined by the softMax classification layer and compared with the preset label. The error is used to train the model parameters of the entire quantum hybrid neural network through the BP method. The minimum mean square error loss function and the fastest gradient descent optimization method are used to determine the optimal learning rate and number of iterations based on experiments.

[0062] The one-dimensional image test sample of the target is input into the trained quantum hybrid neural network, and the corresponding entanglement features are extracted as the input of the softMax classification layer. The label corresponding to the maximum component in the output vector of the softMax classification layer is used as the target recognition category.

[0063] Simulation Example

[0064] The data set used in this simulation experiment is the simulated HRRP sample data of AH64 and F15 aircraft obtained through electromagnetic simulation calculation. The radar carrier frequency in the simulation calculation is 6GHz, the bandwidth is 400MHz, the pitch angle of the target is 3 degrees, and the HRRP data is collected at intervals of 0.1° in the azimuth range of 0° to 180°. Finally, 1801 HRRPs were collected for each type of aircraft, of which the number of range units was 320. From each type of target with an azimuth angle of 0° to 180°, 90 HRRP samples were extracted at intervals of 1° for training, and 1711 HRRP samples were used for testing.

[0065] Noise with a signal-to-noise ratio of 0 dB is added to the data set, and TARAN, T-SCNN, LSTM and the method in this paper are used to classify these data. The recognition results are shown in Table 1. Among them, the recognition result is the average recognition rate of 10 repeated experiments.

[0066] Table 1 Recognition rate statistics (SNR is 0dB)

[0067] method CNN-QC TARAN T-SCNN LSTM Recognition rate (%) 89 83 84 81

[0068] As can be seen from Table 1, the recognition rate of the CNN-QC method proposed in this paper is 5% higher than that of the T-SCNN method, which shows that the proposed method is effective.

Claims

1. A radar target HRRP recognition method based on quantum hybrid neural network, characterized in that: The following steps are involved: S1, obtain multiple HRRP sample data of multiple categories of targets to build a training data set; S2. Construct a quantum hybrid neural network for target recognition, including a CNN subnet and a quantum subnet. The CNN subnet is used to extract features from the input HRRP sample data to obtain convolution features matching the quantum subnet; The quantum subnet includes a quantum circuit and a classification layer, wherein the quantum circuit includes a first quantum circuit and a second quantum circuit, and the two quantum circuits each include an H gate, a rotating Y gate, a rotating Z gate and a measurement gate connected in sequence, and a CNOT gate shared by the two quantum circuits is provided between the rotating Z gate and the measurement gate; wherein the rotation angles θ of the rotating Y gate and the rotating Z gate are obtained by convolution features; The H gate is used to convert the input quantum bit from the ground state to a superposition state with equal probability, expressed as: Among them, |Q 01 >and|Q 11 > respectively represent the quantum superposition state of the first quantum circuit and the quantum superposition state of the second quantum circuit, |0> and |1> respectively represent the two ground states of quantum; The rotating Y gate is used for the quantum superposition state |Q 01 >and|Q 11 >Perform the operation: i＝0,1 Among them, θ i represents the rotation angle of the rotating Y gate, i=0 means the parameter corresponds to the first quantum circuit, i=1 means the parameter corresponds to the second quantum circuit; The rotating Z gate pairs |Q i2 >Processing yields: Among them, |Q 03 > represents the rotational Z gate output of the first quantum circuit, θ2 is the rotation angle of the rotational Z gate of the first quantum circuit, |Q 13 > represents the rotational Z gate output of the second quantum circuit, θ3 is the rotation angle of the rotational Z gate of the second quantum circuit; The input of the CNOT gate is |Q 03 >and|Q 13 >, the output is: The measurement gate performs quantum measurement on |Q4> and obtains: P ij =<Q4|M ij |Q4> Among them, M ij is the measurement operator, and M ij = |ij><ij|, where |ij> represents the computational ground state, <ij| is the conjugate transpose of |ij>, and i, j = 0, 1; After quantum measurement, the probability value of |Q4> in four calculation bases |00>, |01>, |11>, |10> is obtained, and then input into the classification layer for classification; S3, using the training data set obtained in S1 to train the quantum hybrid neural network constructed in S2, the training process is optimized using the minimum mean square error loss function and the fastest gradient descent optimization method to obtain a trained quantum hybrid neural network; S4. Input the HRRP sample to be identified into the trained quantum hybrid neural network, and use the label corresponding to the maximum component in the output vector of the classification layer as the target identification category.

2. According to claim 1, a radar target HRRP recognition method based on quantum hybrid neural network is characterized in that: The CNN subnet consists of three convolutional layers, a pooling layer and a fully connected layer connected in sequence, wherein the convolutional layer consists of convolution operation, batch normalization and ReLu operation. The size of the convolution kernel in the three convolutional layers is 1*3, and the number of channels is 32, 64 and 64 respectively.

Images