Integrated method of narrowband radar echo denoising and recognition for micro-moving targets in space

Abstract

The present invention discloses an integrated method for narrowband radar echo denoising and identification of space micro-moving targets. The method primarily addresses the issues of suppressed separable features and insufficient utilization of RCS information in existing narrowband radar echo denoising and identification technologies for space micro-moving targets. The method comprises the following steps: acquiring full-angle static electromagnetic echoes of space micro-moving targets to generate a training sample set and a test sample set; constructing an integrated denoising and identification network consisting of a complex-valued attention-enhanced denoising module and a complex-valued temporal convolution recognition module, and constructing a hybrid loss function for the network based on the losses of these two modules; training the integrated network based on the training sample set; and inputting the test sample set into the trained integrated denoising and identification network to obtain a classification result output by the network. The present invention avoids the suppression of separable features during the denoising process, fully utilizes RCS data information, and simultaneously improves denoising and identification performance at low signal-to-noise ratios. The method can be used for identifying space micro-moving targets.

Description

Technical Field

[0001] The present invention belongs to the field of radar technology, and further relates to a radar echo noise reduction and target recognition integration method, which can be used for spatial micro-motion target recognition under low signal-to-noise ratio conditions. Background Art

[0002] "Micromotion" refers to the vibration or rotational state of a target or target component relative to the radar's line of sight, in addition to its main motion, such as spin, precession, and tumble. With the increasing frequency of human exploration of outer space and space activities, a large number of space micromotion targets exist in orbit, posing a significant threat to my country's space security. Existing methods for identifying space micromotion targets can be divided into those based on broadband radar and those based on narrowband radar. Compared to broadband radar, narrowband radar has the advantages of longer range and easier processing. Furthermore, the narrowband RCS of a micromotion target contains important electromagnetic scattering, structural, and motion characteristics. Therefore, RCS-based space micromotion target identification plays a vital role in fields such as space situational awareness. Existing technologies typically involve two separate steps: noise reduction and identification. This causes the noise reduction process to suppress features that are beneficial to identification, resulting in poor recognition performance. Furthermore, existing technologies only utilize the amplitude information of the RCS for noise reduction and identification, without utilizing its phase information. This results in insufficient target information utilization and low recognition rates.

[0003] In their paper "Convolutional neural network for classifying spatial target of the same shape by using RCS timeseries" (IET Radar 2018), Jian Chen, Shiyou Xu, and Zengping Chen proposed a one-dimensional convolutional neural network (RCSnet) to identify the real RCS of a target. Their implementation scheme involves first extracting the real RCS sequence information using a feature extraction network consisting of alternating one-dimensional convolutional and pooling layers; then performing feature extraction and fusion using a fully connected layer; and finally, outputting the target category label using a softmax layer. Because this method does not perform RCS denoising, its recognition performance rapidly degrades under low signal-to-noise ratio conditions. Furthermore, because it only utilizes the amplitude information of the RCS and ignores the separable features contained in the phase, recognition performance degrades or even fails completely.

[0004] In their paper "Radar Target Recognition Method Based on Stacked Denoising Sparse Autoencoders" (Journal of Radar Science, 2017), Zhao Feixiang, Liu Yongxiang, and Huo Kai proposed a radar target recognition method based on a stacked denoising sparse autoencoder. The method's implementation involves first denoising the real RCS data using a stacked denoising sparse autoencoder and extracting hidden features from the stacked denoising sparse autoencoder. These features are then fed into a K-nearest neighbor classifier to output the target category label. Because this method performs denoising and recognition separately, features that are beneficial to recognition are suppressed during the denoising process. Furthermore, since only the amplitude information of the RCS is utilized, the separable features contained in the phase are ignored, limiting recognition performance. Summary of the Invention

[0005] The purpose of the present invention is to address the deficiencies of the above-mentioned prior art and propose an integrated method for narrowband radar echo noise reduction and recognition of spatial micro-motion targets, so as to improve the target recognition performance while improving the noise reduction performance.

[0006] The technical ideas for achieving the purpose of the present invention are: effective denoising of complex RCS sequences is achieved by adopting a complex-valued attention enhancement denoising module; efficient recognition of the denoised RCS sequences is achieved by adopting a complex-valued temporal convolution recognition module; and an integrated denoising and recognition network is designed by constructing a hybrid loss function to integrate denoising loss and recognition loss, thereby avoiding the suppression of separable features during the denoising process and improving target recognition performance.

[0007] According to the above ideas, the implementation steps of the present invention include the following:

[0008] (1) Generate training sample set and test sample set:

[0009] (1a) Establish three-dimensional geometric electromagnetic models of six types of space micro-motion targets, including flat-bottomed cones, spherical-bottomed cones, cone-cylinders, spherical-bottomed cone-cylinders, cone-cylinder combinations, and grooved flat-bottomed cones, and obtain their narrow-band radar cross-section (RCS) data;

[0010] (1b) In the RCS data, the data and corresponding labels of the radar at the initial pitch angle of 31°-40° are selected according to the ratio of 10:3 to form the training sample set, and the data and corresponding labels of the radar at the initial pitch angle of 41°-43° are selected to form the test sample set;

[0011] (2) Constructing an integrated network for noise reduction and recognition:

[0012] (2a) Establish a complex-valued attention-enhanced denoising module consisting of a cascade of complex one-dimensional convolutional layers, a complex attention-enhanced encoder, and a complex one-dimensional deconvolutional layer;

[0013] (2b) Establish a complex-valued temporal convolution recognition module consisting of a three-layer complex temporal convolution module and a cascade of softmax classifiers;

[0014] (2c) Cascading the complex-valued attention enhancement denoising module and the complex-valued temporal convolution recognition module to obtain an integrated denoising and recognition network;

[0015] (3) Construct a hybrid loss function J based on the denoising loss and recognition loss, and use it as the loss function of the denoising and recognition integrated network;

[0016] (4) Using the training sample set, the denoising and recognition integrated network is iteratively trained through the back propagation algorithm until the loss function J converges, and a trained denoising and recognition integrated network is obtained;

[0017] (5) The test sample set is input into the trained integrated denoising and recognition network for testing, and the classification results of the network output are obtained to complete the denoising and recognition of the spatial micro-motion target.

[0018] Compared with the prior art, the present invention has the following advantages:

[0019] First, the present invention integrates noise reduction and recognition losses to construct a hybrid loss function, designing a network that integrates noise reduction and recognition, achieving integrated noise reduction and recognition. Compared to existing techniques that perform noise reduction and recognition separately, the present invention avoids the suppression of separable features during the noise reduction process, thereby improving target recognition performance.

[0020] Second, the complex-valued attention enhancement denoising module and complex-valued temporal convolution recognition module constructed by the present invention fully utilize the data characteristics of complex RCS, and at the same time utilize the amplitude information and phase information of the complex RCS data. Compared with the existing technology that only uses the real-valued network to perform denoising and recognition on the amplitude information of RCS data, the present invention can significantly improve the target recognition rate while improving the denoising performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a general flow chart for realizing the present invention;

[0022] Figure 2 This is the framework diagram of the noise reduction and recognition integrated network in the present invention. DETAILED DESCRIPTION

[0023] The embodiments and effects of the present invention are further described in detail below with reference to the accompanying drawings.

[0024] Reference Figure 1 , the implementation steps of this embodiment include the following.

[0025] Step 1: Obtain full-angle static electromagnetic echoes and generate various sample sets.

[0026] (1.1) Establish six target models: flat-bottomed cone, spherical-bottomed cone, cone-cylinder, spherical-bottomed cone-cylinder, cone-cylinder combination, and grooved flat-bottomed cone. In FEKO software, set the radar carrier frequency to 500 MHz, bandwidth to 20 MHz, azimuth observation angle to 0° to 360°, and elevation observation angle to 0°. Select the standard PO method as the calculation method. Run FEKO software to obtain a .ffe file. Use MATLAB to read the .ffe file and obtain the full-angle static electromagnetic echo for each target type.

[0027] (1.2) Extract the full-angle static electromagnetic echo according to the target's micro-motion form and generate the dynamic electromagnetic echo for each type of target:

[0028] Set the target initial pitch angle to 31° to 43°, changing it at 1° intervals. Measure the precession frequency and precession angle at equal intervals at each pitch angle to obtain the pitch angle change curve.

[0029] According to the pitch angle variation curve, the full-angle static echo of each type of target is extracted to obtain the narrow-band radar cross-section (RCS) data of each type of target;

[0030] (1.3) Select the narrowband RCS data and corresponding labels observed by the radar at an initial pitch angle of 31-40° to form a training sample set;

[0031] (1.4) Select the radar narrowband RCS data and corresponding labels at the initial pitch angle of 41-43° to form a test sample set;

[0032] In this example, at each initial pitch angle, 11 precession frequencies and 11 precession angles are taken at equal intervals. Each target type can generate 1573 narrowband RCS data points. The six target types have a total of 9438 RCS data points. The training sample set has a total of 7260 RCS data points, and the test sample set has a total of 2178 RCS data points.

[0033] Step 2: Construct a denoising module with complex-valued attention enhancement.

[0034] (2.1) Set the convolution kernel size to 20, the stride to 10, the number of convolution kernels to 256, the activation function to complex ReLU, and the output to the complex one-dimensional convolution layer with noisy sequence;

[0035] (2.2) Establish a complex attention-enhanced encoder consisting of a complex multi-head attention mechanism, residual connections, complex layer normalization, and a two-layer complex feedforward neural network cascade. The number of multi-head attention heads is 2, the number of neurons in the complex feedforward neural network is 256, and the output is a complex noise mask matrix.

[0036] (2.3) Set the convolution kernel size to 21, the stride to 10, the number of convolution kernels to 1, and output the complex one-dimensional deconvolution layer of the denoised RCS sequence;

[0037] (2.4) The complex one-dimensional convolutional layer, the complex attention-enhanced encoder, and the complex one-dimensional deconvolutional layer are cascaded to form a complex-valued attention-enhanced denoising module, where the input of the complex one-dimensional deconvolutional layer is the Hadamard product of the output of the complex one-dimensional convolutional layer and the output of the complex attention-enhanced encoder.

[0038] Step 3: Construct a complex-valued temporal convolution recognition module.

[0039] (3.1) Establish the first complex temporal convolution module consisting of a cascade of a complex dilated convolution layer, a complex BatchNorm layer, a complex ReLU activation function, a dropout layer, and a residual connection layer. The number of convolution kernels in the complex dilated convolution layer is 32, the convolution kernel size is 11, the dilation rate is 1, and the retention probability of the dropout layer is 0.9.

[0040] (3.2) Establish a second complex temporal convolution module consisting of a cascade of a complex dilated convolution layer, a complex BatchNorm layer, a complex ReLU activation function, a dropout layer, and a residual connection layer. The number of convolution kernels in the complex dilated convolution layer is 64, the convolution kernel size is 11, the dilation rate is 4, and the retention probability of the dropout layer is 0.9.

[0041] (3.3) Establish a third complex temporal convolution module consisting of a cascade of a complex dilated convolution layer, a complex BatchNorm layer, a complex ReLU activation function, a dropout layer, and a residual connection layer. The number of convolution kernels in the complex dilated convolution layer is 128, the convolution kernel size is 11, the dilation rate is 8, and the retention probability of the dropout layer is 0.9.

[0042] (3.4) Build a Softmax classifier with 6 neurons, whose input is the feature vector obtained by flattening the output of the third complex time convolution module in (3.3);

[0043] (3.5) The first complex-time convolution module, the second complex-time convolution module, the third complex-time convolution module, and the Softmax classifier are sequentially cascaded to form a complex-time convolution recognition module.

[0044] Step 4: Design a noise reduction and recognition integrated network.

[0045] The complex-valued attention-enhanced denoising module constructed in step 2 and the complex-valued temporal convolution recognition module constructed in step 3 are cascaded to form an integrated denoising and recognition network, as shown in Figure 2 shown.

[0046] Step 5: Set the loss function of the denoising and recognition integrated network.

[0047] (5.1) Set the denoising loss L of the complex-valued attention-enhanced denoising module den for:

[0048]

[0049] in, represents expectation, Re· and Im represent the real and imaginary parts of the complex number, respectively, and x clean represents clean complex RCS data without noise, x denoise Represents the complex RCS data after noise reduction;

[0050] (5.2) Set the recognition loss L of the complex-valued temporal convolution recognition module recog for:

[0051]

[0052] Among them, y k is the kth element of the true category label of the sample, Represents the kth element of the sample prediction label, where K represents the total number of target categories;

[0053] (5.3) According to the noise reduction loss L den and recognition loss L recog Construct the hybrid loss function J:

[0054] J=α*L den +β*L recog

[0055] Where α is the noise reduction loss coefficient, β is the recognition loss coefficient, α = β = 0.5;

[0056] (5.4) The hybrid loss function J is used as the loss function of the denoising and recognition integrated network.

[0057] Step 6: Input the training sample set into the denoising and recognition integrated network, and train it through the back propagation algorithm to obtain a trained denoising and recognition integrated network.

[0058] (6.1) Initialize the learnable parameters w in the denoising and recognition integrated network;

[0059] (6.2) Input the training sample set into the complex-valued attention enhancement and denoising module, and perform calculations layer by layer along the structure of the complex-valued attention enhancement and denoising module to obtain the denoised complex RCS sequence;

[0060] (6.3) The denoised complex RCS sequence is input into the complex-valued time convolution recognition module, and calculations are performed layer by layer along the structure of the complex-valued time convolution recognition module to obtain the classification result;

[0061] (6.4) Based on the denoised complex RCS sequence obtained in (6.2) and the classification result obtained in (6.3), calculate the value of the mixing function J according to the mixing function formula in step 5;

[0062] (6.5) The learnable parameter w in the denoising and recognition integrated network is updated according to the following formula:

[0063]

[0064] in, is the updated network parameter, w is the network parameter before the update, η is the learning rate, is the gradient of the loss function J with respect to w;

[0065] (6.6) Using the updated network parameters Repeat (6.2) to (6.5) for iterative update until the hybrid loss function J converges and stops iteration to obtain a trained denoising and recognition integrated network.

[0066] In step 7, the generated test sample set is input into the trained integrated noise reduction and recognition network for testing, and the classification result of the network output is obtained to complete the noise reduction and recognition of the spatial micro-motion target.

[0067] The effects of the present invention are further described below in conjunction with simulation experiments.

[0068] 1. Simulation experiment conditions:

[0069] The simulation experiment of the present invention uses the standard PO method to generate narrowband radar echo data of six types of spatial micro-motion targets, which are: flat-bottomed cone, spherical-bottomed cone, cone-cylinder, spherical-bottomed cone-cylinder, cone-cylinder combination and grooved flat-bottomed cone.

[0070] In the FEKO software, the radar carrier frequency is set to 500 MHz and the narrowband signal bandwidth is set to 20 MHz. When extracting narrowband RCS data, the pulse repetition frequency is set to 20 Hz, the observation time is set to 10 s, and the initial pitch angle of each type of target is set to 31°-43°, with an interval of 1°. At each pitch angle, 11 precession frequencies and 11 precession angles are extracted at equal intervals.

[0071] Each type of target generates 1573 narrowband RCS data, and the six types of targets generate a total of 9438 narrowband RCS data.

[0072] The samples and labels corresponding to initial pitch angles of 31° to 40° were used as the training set, resulting in 7260 training samples. The samples and labels corresponding to initial pitch angles of 41° to 43° were used as the test set, resulting in 2178 test samples. Gaussian white noise was added to the training and test sample sets to generate training and test sample sets with signal-to-noise ratios of 0dB, 5dB, and 10dB, respectively.

[0073] The hardware platform for the simulation experiment is Intel(R)Core(TM)i7-10700@2.90GHz CPU and 16GB RAM;

[0074] The simulation experiment software platform is MATLAB 2018b and Python 3.6.

[0075] 2. Simulation experiment content and result analysis:

[0076] Simulation experiment: Under the above simulation conditions, based on the training sample sets and test sample sets with different signal-to-noise ratios, the existing radar target recognition method based on stacked denoising sparse autoencoder and the method of the present invention are respectively used to classify and recognize the radar narrowband echo data of 6 types of spatial micro-motion targets under different signal-to-noise ratios, and the respective recognition rates c and root mean square errors RMSE are calculated as evaluation indicators. The results are shown in Table 1.

[0077] Table 1 Comparison of recognition rate and RMSE between existing methods and the present invention at various signal-to-noise ratios

[0078]

[0079] The calculation formulas for the recognition rate c and root mean square error RMSE in Table 1 are as follows:

[0080]

[0081]

[0082] Where M represents the number of sequence samples in the test sample set, h(·) represents the classification and discrimination function, and t i represents the true category of the i-th test sample in the test sample set, y i Indicates the network output result corresponding to the i-th test sample in the test sample set. When t i and y i When they are equal, h(t i ,y i ) is equal to 1, otherwise h(t i ,y i ) is equal to 0; represents the RCS sample after denoising of the i-th image, x irepresents the i-th clean, noise-free RCS sample, j represents the j-th element in the i-th RCS sample, N represents the dimension of the i-th RCS sample, and T represents the total number of test samples. In this simulation, N = 201 and T = 2178. A smaller RMSE value indicates that the denoised RCS sample is closer to the clean, noise-free RCS sample, indicating better noise reduction performance.

[0083] It can be seen from Table 1 that under the signal-to-noise ratios of 0dB, 5dB, and 10dB, the noise reduction performance and recognition performance of the present invention are significantly improved compared with the prior art.

[0084] Comprehensive simulation experiments show that the integrated narrowband radar echo denoising and recognition method of the present invention can avoid the suppression of separable features during the denoising process and can simultaneously utilize the amplitude and phase information of the complex RCS data, thereby improving both the denoising performance and the target recognition performance.

Claims

1. A method for integrating narrowband radar echo noise reduction and recognition of spatial micro-motion targets, characterized in that: The steps include: (1) Generate training sample set and test sample set: (1a) Establish three-dimensional geometric electromagnetic models of six types of space micro-motion targets, including flat-bottomed cones, spherical-bottomed cones, cone-cylinders, spherical-bottomed cone-cylinders, cone-cylinder combinations, and grooved flat-bottomed cones, and obtain their narrow-band radar cross-section (RCS) data; (1b) In the RCS data, the data and corresponding labels of the radar at the initial pitch angle of 31°-40° are selected according to the ratio of 10:3 to form the training sample set, and the data and corresponding labels of the radar at the initial pitch angle of 41°-43° are selected to form the test sample set; (2) Constructing an integrated network for noise reduction and recognition: (2a) Establish a complex-valued attention-enhanced denoising module consisting of a cascade of complex one-dimensional convolutional layers, a complex attention-enhanced encoder, and a complex one-dimensional deconvolutional layer; (2b) Establish a complex-valued temporal convolution recognition module consisting of a three-layer complex temporal convolution module and a cascade of softmax classifiers; (2c) Cascading the complex-valued attention enhancement denoising module and the complex-valued temporal convolution recognition module to obtain an integrated denoising and recognition network; (3) Construct a hybrid loss function J based on the denoising loss and recognition loss, and use it as the loss function of the denoising and recognition integrated network; (4) Using the training sample set, the denoising and recognition integrated network is iteratively trained through the back propagation algorithm until the loss function J converges, and a trained denoising and recognition integrated network is obtained; (5) The test sample set is input into the trained integrated denoising and recognition network for testing, and the classification results of the network output are obtained to complete the denoising and recognition of the spatial micro-motion target.

2. The method according to claim 1, wherein: In step (1a), the narrowband radar cross-section (RCS) data of the three-dimensional geometric electromagnetic model of the six types of space micro-motion targets are obtained as follows: (1a1) In the electromagnetic simulation software FEKO, set the radar carrier frequency to 500 MHz, the narrowband signal bandwidth to 20 MHz, the azimuth observation angle to 0°, and the elevation observation angle to 0°-360°; (1a2) Using the standard PO method, run FEKO software to obtain .ffe files, and use MATLAB program to read the .ffe files to obtain the full-angle static electromagnetic echo of each type of target; (1a3) The radar's initial pitch angle is set to 31°-43°. The precession frequency and precession angle of the micro-motion target are varied at equal intervals to obtain a pitch angle variation curve. Based on the pitch angle variation curve, the full-angle static echo of each type of target is extracted to obtain the narrowband radar cross-section (RCS) data of each type of target at the initial pitch angle.

3. The method according to claim 1, wherein: The complex one-dimensional convolutional layer, the complex attention-enhanced encoder, and the complex one-dimensional deconvolutional layer that constitute the complex-valued attention-enhanced denoising module in step (2a) have the following structural parameters: The complex one-dimensional convolution layer has 256 convolution kernels, a convolution kernel size of 20, a stride of 10, and outputs a noisy sequence; The complex one-dimensional deconvolution layer has a convolution kernel number of 1, a convolution kernel size of 21, a stride of 10, and outputs a denoised RCS sequence; The multiple attention-enhanced encoder consists of a complex multi-head attention mechanism, complex layer normalization, a two-layer complex feedforward neural network, and a complex layer normalization cascade, and each layer has a residual connection before normalization. The number of multi-head attention heads is 2, the number of neurons in the two-layer complex feedforward neural network is 256, and the output is a complex noise mask matrix; The noisy sequence output by the complex one-dimensional convolution layer is multiplied by the complex noise mask matrix output by the complex attention-enhanced encoder, and then input into the complex one-dimensional deconvolution layer to obtain the denoised RCS sequence.

4. The method according to claim 1, wherein: The structural parameters of the three-layer complex-valued temporal convolution module and the Softmax classifier in step (2b) are as follows: The first layer of complex temporal convolution module is composed of a cascade of complex dilated convolution layer, complex BatchNorm layer, complex ReLU activation function, and dropout layer. The number of convolution kernels in the complex dilated convolution layer is 32, the convolution kernel size is 11, the void ratio is 1, and the retention probability of the dropout layer is 0.

9. The second layer of complex temporal convolution module is composed of a cascade of complex dilated convolution layer, complex BatchNorm layer, complex ReLU activation function, and dropout layer. The number of convolution kernels in the complex dilated convolution layer is 64, the convolution kernel size is 11, the void ratio is 4, and the retention probability of the dropout layer is 0.

9. The third layer of complex temporal convolution module is composed of a cascade of complex dilated convolution layer, complex BatchNorm layer, complex ReLU activation function, and dropout layer. The number of convolution kernels in the complex dilated convolution layer is 128, the convolution kernel size is 11, the void ratio is 8, and the retention probability of the dropout layer is 0.

9. The number of neurons in the Softmax classifier is 6.

5. The method according to claim 1, wherein: The hybrid loss function J constructed in step (3) is expressed as follows: J=α*L den +β*L recog Among them, α is the noise reduction loss coefficient, β is the recognition loss coefficient, α=β=0.5, L den is the noise reduction loss, L rec o g Identification loss, these two losses are expressed as follows: in, represents expectation, Re· and Im· represent the real part and imaginary part of the complex number respectively, and x clean represents the clean complex radar cross section RCS, x denoise represents the complex RCS after noise reduction, y k is the kth element of the true category label of the sample, Represents the kth element of the sample prediction label, and K represents the total number of target categories.

6. The method according to claim 1, wherein: Step (5) uses the training sample set to iteratively train the noise reduction and recognition integrated network through the back propagation algorithm, and is implemented as follows: (6a) Initialize the learnable parameters w in the denoising and recognition integrated network; (6b) Input the training sample set into the complex-valued attention enhancement and denoising module, and perform calculations layer by layer along the structure of the complex-valued attention enhancement and denoising module to obtain the denoised complex RCS sequence; (6c) Inputting the denoised complex RCS sequence into the complex-valued time convolution recognition module, performing calculations layer by layer along the structure of the complex-valued time convolution recognition module, and obtaining the classification result; (6d) Calculating the value of the mixing function J according to the mixing function formula in step 5 based on the denoised complex RCS sequence obtained in (6b) and the classification result obtained in (6c); (6e) The learnable parameter w in the denoising and recognition integrated network is updated according to the following formula: in, is the updated network parameter, w is the network parameter before the update, η is the learning rate, is the gradient of the loss function J with respect to w; (6f) Using the updated network parameters Repeat (6b) to (6e) for iterative update until the hybrid loss function J converges and stops iteration to obtain a trained denoising and recognition integrated network.

Images