A deep learning time-frequency representation method for radar micro-motion features

Abstract

The present invention relates to a deep learning time-frequency representation method for radar micro-motion features, and belongs to the field of radar signal processing technology. The present invention first pre-processes radar echo data to obtain target information data containing a distance dimension; then designs a deep learning network model that represents the micro-motion features of radar targets, and the model includes four parts: a basis function selection module, a feature aggregation module, an energy concentration module, and an output module. A data set is constructed to train and verify the model; finally, the pre-processed measured target data is input to complete the time-frequency representation of the target micro-motion features based on a deep learning method. The present invention can achieve high-resolution micro-motion feature representation of radar targets. Compared with traditional time-frequency analysis methods, it not only shows the advantage of a high degree of energy aggregation, but also shows the details of micro-motion features that traditional time-frequency analysis methods fail to show, thereby improving the representation capability of the target's micro-motion features.

Description

Technical Field

[0001] The present invention relates to a deep learning time-frequency representation method for radar micro-motion features, and belongs to the technical field of radar signal processing. Background Art

[0002] Drones and flying birds are typical "low, slow and small" targets, with low observable properties such as low flight altitude, strong maneuverability and small radar cross-sectional area, making it difficult for radar to detect and identify them. In recent years, with the development of drones, various illegal flying incidents have also emerged one after another, seriously affecting flight safety. Therefore, the detection and identification of "low, slow and small" targets is of great significance.

[0003] The micro-Doppler effect is a phenomenon that when the radar irradiates a moving target, in addition to the overall translation of the target, there may also be small movements such as rotation and vibration, forming additional sidebands on the Doppler spectrum. The analysis and extraction of micro-motion characteristics has become an important research aspect of radar target classification, detection and recognition. When analyzing the micro-Doppler characteristics of a target, the signal is usually transformed into the time-frequency domain. The micro-motion target signal has the characteristics of non-stationarity and time-varying. Analysis in the time-frequency domain can well reveal the frequency components and instantaneous change characteristics contained in the signal. The micro-motion radar echo of the target is usually a multi-component signal, and the signal component is relatively weak relative to the main echo intensity. Therefore, it is particularly important to study the time-frequency distribution with high time-frequency resolution to characterize the micro-motion characteristics of the target.

[0004] In recent years, many scholars have proposed to obtain the micro-motion characteristics of target signals by traditional time-frequency analysis methods. However, most of the methods proposed by scholars to analyze micro-motion characteristics are post-processing methods based on short-time Fourier transform (STFT). The effect of the analysis is largely limited by factors such as the STFT window length, which ultimately leads to unsatisfactory analysis of target micro-motion characteristics. With the development of artificial intelligence, deep learning has been widely used in the field of signal processing. The deep learning method has a strong generalization ability and can use data to drive adaptive learning. However, when the deep learning method is used for the time-frequency representation of micro-Doppler characteristics, sometimes wrong representations will occur, making it difficult to show the micro-motion characteristics of the target and produce accurate and high-resolution time-frequency images. The micro-motion characteristics are the characteristics of small vibrations and attitude changes of rotorcraft drones during flight. In-depth research on the representation of micro-motion characteristics is of great significance for the classification and recognition of "low, slow and small" targets. Therefore, designing a suitable time-frequency method based on deep learning to represent the micro-motion characteristics of targets plays an important role in the recognition and classification of targets. Summary of the invention

[0005] The purpose of this invention is to address the serious time-frequency energy diffusion and insufficient micro-motion feature representation capabilities of traditional time-frequency analysis methods in representing micro-Doppler features, and propose a deep learning time-frequency representation method for radar micro-motion features, aiming to improve the micro-Doppler feature representation capabilities of radar targets.

[0006] The present invention provides a radar micro-motion feature deep learning time-frequency representation method, which is characterized by comprising the following steps:

[0007] Step 1: Data preprocessing:

[0008] The radar collects the original echo signal, performs data preprocessing on the original echo signal, and then performs pulse compression, intercepts the target distance range and selects the target distance unit in sequence to obtain a one-dimensional complex value signal containing the target unit;

[0009] Step 2: Build the ETF-MDNet network model:

[0010] Design an ETF-MDNet network model for deep learning time-frequency representation of one-dimensional complex-valued signals containing target units, including a signal input module, a basis function selection module, a feature aggregation module, and an energy concentration module;

[0011] Step 3: Training and verification of ETF-MDNet network model:

[0012] After building the ETF-MDNet network model, set the training data set, validation data set and loss function of the ETF-MDNet network model to train and validate the network model;

[0013] Step 4: Representation of micro-motion characteristics of measured target data:

[0014] The ETF-MDNet network model completes training and verification, and is then loaded for micro-Doppler feature representation of measured target data.

[0015] Preferably, the signal input module in step 2 performs normalization processing on the input one-dimensional complex-valued signal, which is expressed as:

[0016] ;

[0017] in, is the input one-dimensional complex valued signal, is the minimum value of the input, is the maximum value of the input, is the normalized result;

[0018] Preferably, the basis function selection module in step 2 is implemented by a convolution layer of a convolution module in a convolutional neural network, and the convolution process of the convolution layer without a bias term is expressed by the following formula:

[0019] ;

[0020] in, It can represent the instantaneous frequency at time m, represents the mth segment of the input signal, and W represents the scale L W ×N f The convolution kernel weight matrix of C;

[0021] In the time-frequency analysis method, the discrete form of the Fourier transform is:

[0022] ;

[0023] in, N represents the number of points at which the discrete Fourier transform is performed, compared to and It can be seen that with the increase of training times and the update of network weights, when hour, It can be regarded as Fourier transform, that is, the weight matrix of the network will be fitted as the basis function of Fourier transform;

[0024] Preferably, the feature aggregation module in step 2 introduces channel and spatial attention in the attention mechanism to improve the representation ability of the model micro-motion features, and the channel attention mechanism is:

[0025] ;

[0026] in, F is the input feature map, Avgpool and Maxpool represent global average pooling and maximum pooling operations respectively, MLP represents multi-layer perceptron, and σ represents Sigmoid activation function.

[0027] The spatial attention mechanism is:

[0028] ;

[0029] in, It means to concatenate the results of average pooling and maximum pooling along the channel axis. Represents a 7×7 convolution operation;

[0030] Preferably, the energy concentration module in step 2 uses a residual-connected convolutional autoencoder to achieve high aggregation of the feature map after the concentrated time-frequency features are aggregated, and obtain a highly aggregated time-frequency map; in the energy concentration module, there are 20 layers of encoders and 20 layers of decoders connected by a residual structure, each encoder is composed of a two-dimensional convolution layer and a ReLU layer, and each decoder includes a two-dimensional transposed convolution layer and a ReLU layer; the module uses a Padding operation to retain the width and height of the feature map, the feature map from the encoder is added to the feature map from its corresponding mirror decoder through a residual structure connection, and a one-dimensional transposed convolution layer is used to integrate the complementary information of each group of feature maps to achieve twice the resolution of the frequency dimension. After being processed by the energy concentration module, a high-resolution time-frequency map is obtained;

[0031] Preferably, the training data set of the ETF-MDNet model network in step 3 is constructed as follows:

[0032] The ETF-MDNet dataset consists of a 3×10 4 The training set of samples and a 1×10 3 The validation set of the sample, the data set samples are S ( n ), IF p ( n ), the samples are mixed with additive white Gaussian noise, the signal-to-noise ratio is between −5 and 25 dB, and the input of the training set is a multi-component simulation signal, where the simulation signal and the instantaneous frequency are as follows:

[0033] ;

[0034] ;

[0035] Among them, the number of signal components P Obeys uniform distribution U(1,5), Indicates p The strength of the component, It obeys the standard Gaussian distribution. Indicates p The amplitude of the components follows the uniform distribution U(0.01,3), and The initial phase θ Obeys uniform distribution U(0,2π), and They obey uniform distribution U(-150,150), U(-20,20), signal length L and frequency points N respectively f Set to 256 and 128 respectively. After the energy concentration module, the size of the feature map will become 2×L×Nf , In addition, the corresponding ideal two-dimensional time-frequency matrix is:

[0036] ;

[0037] Where n=0,1,…,L-1 and k=0,1,…,2N f -1 represents the time index and frequency index, ∆f= f s / 2N f represents the frequency interval, where is the logarithmic form of intensity, and ,in F s is the sampling frequency.

[0038] Preferably, the loss function in step 3 is defined as:

[0039] ;

[0040] ;

[0041] in, Represents the time-frequency results generated by the ETF-MDNet network model, is a two-dimensional penalty matrix, Represents the square of the logarithmic amplitude of the maximum amplitude component in the training set.

[0042] Preferably, the steps of training and verifying the network in step 3 are as follows:

[0043] The verification of the ETF-MDNet network model selects the simulation signal Signal of two-component sinusoidal frequency modulation:

[0044] ;

[0045] ;

[0046] The results were compared and analyzed after being processed by short-time Fourier transform, generalized Warblet, redistributed spectrum and ETF-MDNet network model.

[0047] Preferably, the measured target data in step 4 selects fixed-wing UAV, model bird and AC311 helicopter data, and inputs them into the trained ETF-MDNet network model to represent micro-motion features.

[0048] The invention provides a radar micro-motion feature deep learning time-frequency representation method, which has the following beneficial effects:

[0049] (1) The network designed by this method for extracting micro-motion features of radar targets (ETF-MDNet network model) has improved the micro-Doppler feature representation capability of targets such as flying birds and drones, as shown by actual analysis results.

[0050] (2) Compared with the traditional STFT, GWoRblet and RS time-frequency analysis methods, the Rényi entropy value of this method is lower, indicating that the time-frequency energy is more concentrated. At the same time, it shows the details of micro-motion features that are not shown by traditional methods, demonstrating the effectiveness of this method and proving that the ETF-MDNet network model has more obvious effects in energy concentration and micro-motion feature representation. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a flow chart of the present invention;

[0052] Figure 2 This is the structure diagram of the ETF-MDNet network model;

[0053] Figure 3 Select the block diagram for the basis function;

[0054] Figure 4 It is the feature aggregation module diagram;

[0055] Figure 5 The two-component simulation signal network verification result diagram, where (a) is the IF instantaneous frequency result diagram, (b) is the STFT processing result diagram, (c) is the GWorblet processing result diagram, (d) is the RS processing result diagram, and (e) is the ETF-MDNet network model processing result diagram;

[0056] Figure 6 The following are the result diagrams of the micro-motion feature representation of the model flying bird, where (a) is the result diagram of STFT processing, (b) is the result diagram of GWorblet processing, (c) is the result diagram of RS processing, and (d) is the result diagram of ETF-MDNet network model processing. DETAILED DESCRIPTION

[0057] The following will be combined with the attached Figure 1-6 The specific steps of the present invention are described in detail.

[0058] A radar micro-motion feature deep learning time-frequency representation method of this embodiment includes the following steps:

[0059] (1) Collect target echo data and perform data preprocessing.

[0060] The target data collected by the radar is preprocessed, the UAV radar echo data is collected, and the echo data is gradually pulse compressed, the target distance range is intercepted, and the target distance unit is selected to obtain the target data of the one-dimensional complex value signal.

[0061] (2) Design a deep learning time-frequency representation network (ETF-MDNet network model) for analyzing radar target characteristics, which includes a signal input module, a basis function selection module, a feature aggregation module, and an energy concentration module.

[0062] a.Signal input module:

[0063] Combination Figure 2 ,ETF-MDNet network model structure diagram,The input target micro-Doppler signal is a preprocessed complex-valued signal. Since the network model contains nonlinear activation functions, in order to protect the feature map of the early input layer, prevent the feature map from being too large or too small and unable to enter the saturation area and accelerate the convergence of the network, the input one-dimensional complex-valued signal must be uniformly normalized.

[0064] ;

[0065] in, is the input one-dimensional complex valued signal, is the minimum value of the input, is the maximum value of the input, is the normalized result.

[0066] b. Basis function selection module:

[0067] Combination Figure 3 , Basis function selection module diagram, the basis function selection module is composed of one-dimensional complex-valued convolution, the convolution process of the convolution module can be expressed by the following formula, using Padding to maintain the signal length, and using complex-valued convolution kernels of specific scales to learn various basis functions. The operation of the convolution layer without bias terms can be described as follows:

[0068] ;

[0069] in, It can represent the instantaneous frequency at time m, represents the mth segment of the input signal, and W represents the scale L W ×N f C is the convolution kernel weight matrix.

[0070] In the traditional time-frequency analysis method, the discrete form of Fourier transform is:

[0071] ;

[0072] Where N represents the number of points for discrete Fourier transform. and It can be seen that with the increase of training times and the update of network weights, when hour, It can be regarded as a general Fourier transform, that is, the weight matrix of the network is fitted as the basis function of the Fourier transform.

[0073] However, relying solely on a set of Fourier basis functions cannot achieve sufficient time-frequency energy aggregation. In order to further fully characterize and express the high-resolution micro-Doppler characteristics of the target, we set the ETF-MDNet network model to learn the basis functions of group C to improve the aggregation of time-frequency energy, that is, we can obtain the feature map of group C.

[0074] c. Feature aggregation module:

[0075] Combination Figure 4 ,Feature aggregation module diagram,The feature aggregation module is mainly implemented through the attention mechanism. According to the micro-motion characteristics of radar targets, the channel attention (Channel Attention) and spatial attention (Spatial Attention) in the attention mechanism are introduced to enhance the ETF-MDNet network model's ability to represent micro-motion features. By adaptively optimizing various complete basis functions generated by data drive in the channel and spatial dimensions, a feature map of aggregated micro-Doppler features is formed. This dual attention mechanism enables the model to adaptively focus on the key channels and spatial positions in the input features, achieving key attention to the target micro-motion information and suppression of irrelevant noise.

[0076] Channel Attention:

[0077] ;

[0078] Among them, F is the input feature map, Avgpool and Maxpool represent global average pooling and maximum pooling operations respectively, MLP represents multi-layer perceptron, and σ represents Sigmoid activation function.

[0079] Spatial Attention:

[0080] ;

[0081] in, It means to concatenate the results of average pooling and maximum pooling along the channel axis. Represents a 7×7 convolution operation.

[0082] d. Energy concentration module:

[0083] Residual Encoder Decoder blocks are used in this module to concentrate the energy of feature maps and obtain high-resolution micro-motion feature maps. In this module, there are 20 layers of encoders and 20 layers of decoders connected by residual structures. Each encoder consists of a 2D convolution layer and a ReLU layer, and each decoder includes a 2D transposed convolution layer and a ReLU layer. This module also uses Padding operations to preserve the width and height of the feature map. During use, the feature map from the encoder is added to the feature map from its corresponding mirror decoder through the residual structure connection.

[0084] Finally, a one-dimensional transposed convolution layer is used to integrate the complementary information of each group of feature maps to achieve double resolution in the frequency dimension. After processing by the energy concentration module, a high-resolution time-frequency map can be obtained.

[0085] (3) Construction of the ETF-MDNet network model training set and setting of the loss function, as well as network training and verification.

[0086] a. Construction of ETF-MDNe network model training set:

[0087] The dataset of the ETF-MDNet network model includes a 3×10 4 The training set consists of 1×10 3 The validation set of samples. The following three data sets are generated. The samples are mixed with additive white Gaussian noise, and the signal-to-noise ratio is between −5 and 25 dB.

[0088] The input of the training set is a multi-component simulation signal, where the simulation signal and the instantaneous frequency (IF) are as follows:

[0089] ;

[0090] ;

[0091] Among them, the number of signal components P follows the uniform distribution U (1,5), represents the intensity of the pth component, It obeys the standard Gaussian distribution. represents the amplitude of the pth component, which follows the uniform distribution U(0.01,3). and The initial phase θ Obeys uniform distribution U(0,2π), and They obey uniform distribution U(-150,150) and U(-20,20) respectively. Signal length L and frequency points N f Set to 256 and 128 respectively. After the energy concentration module, the size of the feature map will become 2×L×N f . In addition, the corresponding ideal two-dimensional time-frequency matrix is:

[0092] ;

[0093] Where n=0,1,…,L-1 and k=0,1,…,2N f -1 represents the time index and frequency index, ∆f= f s / 2N f represents the frequency interval, where is the logarithmic form of intensity, and ,in F s is the sampling frequency.

[0094] b. Setting of loss function:

[0095] The loss function is defined as:

[0096] ;

[0097] ;

[0098] in, represents the time-frequency results generated by the network ETF-MDNet network model, is a two-dimensional penalty matrix, Represents the square of the logarithmic amplitude of the maximum amplitude component in the training set.

[0099] c. Training and verification of ETF-MDNet model:

[0100] The verification of the ETF-MDNet network model selects the simulation signal Signal of two-component sinusoidal frequency modulation:

[0101] ;

[0102] ;

[0103] Combination Figure 5As shown in the figure of the two-component simulation signal network verification result, the simulation data were processed by short-time Fourier transform (STFT), generalized Warblet (GWarblet), redistributed spectrogram (RS) and ETF-MDNet network model respectively. The comparison of the results shows that the proposed ETF-MDNet network model method can well display the time-frequency analysis results of the signal, while showing the energy aggregation that traditional methods do not have, and can better show the time-frequency trajectory change characteristics of the component intersection points.

[0104] (4) Time-frequency representation of micro-motion features by deep learning of measured target data:

[0105] The signal preprocessed by (1) is input into the trained ETF-MDNet network model to perform time-frequency representation of micro-Doppler features based on deep learning. The fixed-wing UAV, model bird and AC311 helicopter data are selected respectively. Taking the measured model bird data as an example, Figure 6 As shown in the result diagram of the micro-motion feature representation of the model flying bird, the STFT method is insufficient in terms of energy concentration and time-frequency resolution. The results of the GWorblet transform processing lack the continuity of the Doppler frequency and the concentration of the time-frequency energy. This will lead to insufficient detail capture and affect the understanding of the motion state of the model bird. The RS analysis method introduces a large number of irrelevant components in the processing process. These components are irrelevant to the micro-motion features and will interfere with the accuracy of the analysis results. In contrast, the ETF-MDNet network model processing method proposed in this method can continuously display the motion state of the model bird's flapping wings, and effectively suppress other components that are not related to the micro-motion features, showing significant advantages in the processing results. This method not only excels in energy concentration, but also is more refined in depicting micro-motion features.

[0106] In summary, the processing results combined with the measured target data show that this method not only performs well in energy concentration, but also reaches a high level in visual resolution, and comprehensively displays the micro-motion characteristics of radar targets. This can not only improve the accuracy of micro-motion feature parameter estimation results in the future, but also provide more abundant information for target recognition and classification.

[0107] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is determined by the attached claims.

Claims

1. A deep learning time-frequency representation method for radar micro-motion features, characterized in that The following steps are involved: Step 1: Data preprocessing: The radar collects the original echo signal, performs data preprocessing on the original echo signal, and then performs pulse compression, intercepts the target distance range and selects the target distance unit in sequence to obtain a one-dimensional complex value signal containing the target unit; Step 2: Build the ETF-MDNet network model: Design an ETF-MDNet network model for deep learning time-frequency representation of one-dimensional complex-valued signals containing target units, including a signal input module, a basis function selection module, a feature aggregation module, and an energy concentration module; Step 3: Training and verification of ETF-MDNet network model: After building the ETF-MDNet network model, set the training data set, validation data set and loss function of the ETF-MDNet network model to train and validate the network model; Step 4: Representation of micro-motion characteristics of measured target data: The ETF-MDNet network model completes training and verification, and is then loaded for micro-Doppler feature representation of measured target data.

2. A radar micro-motion feature deep learning time-frequency representation method according to claim 1, characterized in that The signal input module performs normalization processing on the input one-dimensional complex valued signal, which is expressed as: ； in, is the input one-dimensional complex valued signal, is the minimum value of the input, is the maximum value of the input, is the normalized result.

3. A radar micro-motion feature deep learning time-frequency representation method according to claim 1, characterized in that The basis function selection module in step 2 is implemented by the convolution layer of the convolution module in the convolutional neural network. The convolution process of the convolution layer without the bias term is expressed by the following formula: ； in, It can represent the instantaneous frequency at time m, represents the mth segment of the input signal, and W represents the scale L W ×N f The convolution kernel weight matrix of C; In the time-frequency analysis method, the discrete form of the Fourier transform is: ； in, N represents the number of points at which the discrete Fourier transform is performed, compared to and It can be seen that with the increase of training times and the update of network weights, when hour, It can be regarded as a Fourier transform, that is, the weight matrix of the network will be fitted as the basis function of the Fourier transform.

4. A radar micro-motion feature deep learning time-frequency representation method according to claim 1, characterized in that The feature aggregation module described in step 2 introduces channel and spatial attention in the attention mechanism to improve the representation ability of the model's micro-motion features. The channel attention mechanism is: ； in, F is the input feature map, Avgpool and Maxpool represent global average pooling and maximum pooling operations respectively, MLP represents multi-layer perceptron, and σ represents Sigmoid activation function; The spatial attention mechanism is: ； in, It means to concatenate the results of average pooling and maximum pooling along the channel axis. Represents a 7×7 convolution operation.

5. A radar micro-motion feature deep learning time-frequency representation method according to claim 1, characterized in that The energy concentration module in step 2 uses a residual-connected convolutional autoencoder to achieve high aggregation of the feature graph after the concentrated time-frequency features are aggregated, thereby obtaining a highly aggregated time-frequency graph; In the energy concentration module, there are 20 layers of encoders and 20 layers of decoders connected by a residual structure. Each encoder consists of a 2D convolution layer and a ReLU layer, and each decoder includes a 2D transposed convolution layer and a ReLU layer. This module uses the Padding operation to retain the width and height of the feature map. The feature map from the encoder is added to the feature map from its corresponding mirror decoder through a residual structure connection. A one-dimensional transposed convolution layer is used to integrate the complementary information of each group of feature maps to achieve twice the resolution of the frequency dimension. After processing by the energy concentration module, a high-resolution time-frequency map is obtained.

6. A radar micro-motion feature deep learning time-frequency representation method according to claim 1, characterized in that The training data set of the ETF-MDNet model network described in step 3 is constructed as follows: The ETF-MDNet dataset consists of a 3×10 4 The training set of samples and a 1×10 3 The validation set of the sample, the data set samples are S ( n ), IF p ( n ), the samples are mixed with additive white Gaussian noise, the signal-to-noise ratio is between −5 and 25 dB, and the input of the training set is a multi-component simulation signal, where the simulation signal and the instantaneous frequency are as follows: ； ； Among them, the number of signal components P Obeys uniform distribution U(1,5), Indicates p The strength of the component, It obeys the standard Gaussian distribution. Indicates p The amplitude of the components follows the uniform distribution U(0.01,3), and The initial phase θ Obeys uniform distribution U(0,2π), and They obey uniform distribution U(-150,150), U(-20,20), signal length L and frequency points N respectively f Set to 256 and 128 respectively. After the energy concentration module, the size of the feature map will become 2×L×N f , In addition, the corresponding ideal two-dimensional time-frequency matrix is: ； Where n=0,1,…,L-1 and k=0,1,…,2N f -1 represents the time index and frequency index, ∆f= f s / 2N f represents the frequency interval, where is the logarithmic form of intensity, and ,in F s is the sampling frequency.

7. A radar micro-motion feature deep learning time-frequency representation method according to claim 6, characterized in that The loss function in step 3 is defined as: ； ； in, Represents the time-frequency results generated by the ETF-MDNet network model, is a two-dimensional penalty matrix, Represents the square of the logarithmic amplitude of the maximum amplitude component in the training set.

8. A radar micro-motion feature deep learning time-frequency representation method according to claim 7, characterized in that The steps for training and validating the network described in step 3 are as follows: The verification of the ETF-MDNet network model selects the simulation signal Signal of two-component sinusoidal frequency modulation: ； ； The results were compared and analyzed after being processed by short-time Fourier transform, generalized Warblet, redistributed spectrum and ETF-MDNet network model.

9. A radar micro-motion feature deep learning time-frequency representation method according to claim 1, characterized in that The measured target data described in step 4, including fixed-wing UAV, model bird and AC311 helicopter data, are input into the trained ETF-MDNet network model to represent micro-motion features.

Images