Micro-motion curve extraction method and system for micro-motion signal time-frequency diagram
The mask extraction model constructed using a self-attention network and residual connection scheme solves the problems of high computational requirements and poor accuracy in the extraction of micro-motion curves from time-frequency diagrams of micro-motion signals in existing technologies, and achieves high reliability and high accuracy in micro-motion curve extraction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CENT SOUTH UNIV
- Filing Date
- 2024-06-06
- Publication Date
- 2026-07-31
AI Technical Summary
Existing methods for extracting micro-motion curves from time-frequency maps of micro-motion signals require knowledge of certain structural parameters and kinematic information of the target, have high computational requirements, and suffer from poor accuracy and reliability when noise and overlapping regions are present, making end-to-end processing impossible.
A mask extraction model based on a self-attention network is adopted. The mask extraction model of the time-frequency map of the micro-motion signal is constructed by using a self-attention network and a residual connection scheme. The micro-motion curve is extracted by using a self-attention module and a residual connection module. Post-processing is performed by combining the Viterbi algorithm and the Fourier series least squares method.
It achieves highly reliable and accurate extraction of micro-motion curves from time-frequency diagrams of micro-motion signals, reducing computational complexity and manual intervention, and improving processing capabilities in noisy and overlapping regions.
Smart Images
Figure CN118734053B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of digital signal processing, and specifically relates to a method and system for extracting micro-motion curves from the time-frequency diagram of a micro-motion signal. Background Technology
[0002] Space cone target identification is a crucial step in missile attack and defense. Because space cone targets need to maintain stability during atmospheric reentry, they exhibit significant precession characteristics during the free-fall phase. This micro-motion characteristic becomes a key feature for space cone target identification. Time-frequency analysis of radar signals can yield a target micro-motion time-frequency map, where the micro-motion curves of the target's scattering points reflect the target's structural and precession information. Extracting the target micro-motion time-frequency map and separating the micro-motion curves of each scattering point is key to target classification.
[0003] Existing methods often require prior knowledge of certain structural parameters and kinematic information of the target. These methods are represented by parametric approaches, including the Rondon transform and the extended Hough transform; they treat the separation of the micro-motion curves as a parameter estimation challenge. While effective, these methods require a very deep understanding of signal modulation, and their reliance on sinusoidal modulation assumptions limits their applicability in specific scenarios. Furthermore, the computational demands increase significantly with the introduction of more parameters.
[0004] In contrast, nonparametric methods are often favored due to their adaptability and applicability to various situations. Common methods include identifying energy peaks or determining the minimum value of the energy functional, but these methods only consider amplitude and frequency variations, resulting in poor accuracy and reliability. Some studies treat the time-frequency representation curve as the target trajectory and use filters such as Kalman filters, modified Kalman filters, or variational Kalman filters to separate the micro-motion curves; however, crosstalk or overlapping regions between instantaneous frequencies can lead to inaccurate tracking. Currently, other methods are designed to mitigate the effects of overlapping regions, such as Ridge Path Reconstruction (RPRG) and RANSAC-based instantaneous frequency estimation algorithms. However, these algorithms are susceptible to noise, and they face serious challenges in cases of significant component overlap, discontinuities, or high noise interference. Furthermore, when the number of components is uncertain, these methods often require manual intervention, such as parameter tuning and iterative control, which can be extremely impractical for processing large amounts of data. In addition, when there are signal interruptions, these methods require additional correlation matching steps, making end-to-end processing impossible. Summary of the Invention
[0005] One of the objectives of this invention is to provide a method for extracting micro-motion curves from time-frequency diagrams of micro-motion signals that is highly reliable, accurate, and efficient.
[0006] The second objective of this invention is to provide a system for extracting the micro-motion curve from the time-frequency diagram of the micro-motion signal.
[0007] The method for extracting the micro-motion curve from the time-frequency diagram of a micro-motion signal provided by this invention includes the following steps:
[0008] S1. Obtain the existing time-frequency diagram of the micro-motion signal and the corresponding micro-motion curve data, and construct a training dataset;
[0009] S2. Based on a self-attention network, a preliminary model for mask extraction of time-frequency maps of micro-motion signals is constructed;
[0010] S3. Using the training dataset constructed in step S1, train the preliminary model for mask extraction of the time-frequency map of the micro-motion signal constructed in step S2 to obtain the mask extraction model of the time-frequency map of the micro-motion signal.
[0011] S4. Input the actual acquired micro-motion signal time-frequency map into the mask extraction model of the micro-motion signal time-frequency map obtained in step S3, and extract the mask information of the target micro-motion signal time-frequency map;
[0012] S5. Post-process the mask information of the target micro-motion signal time-frequency diagram obtained in step S4 to obtain the corresponding micro-motion curve, thus completing the extraction of the micro-motion curve of the target micro-motion signal time-frequency diagram.
[0013] Step S2, which involves constructing a preliminary model for micro-motion curve extraction based on a self-attention network, includes the following steps:
[0014] A preliminary model for extracting micro-motion curves is constructed based on a self-attention network.
[0015] The preliminary model for extracting micro-motion curves includes sub-networks 1 to n, where n is the number of micro-motion curves included in the time-frequency plot of the micro-motion signal.
[0016] Construct the first subnetwork based on the self-attention network;
[0017] Based on self-attention networks and residual connection schemes, subnetworks 2 through n are constructed; the structures of subnetworks 2 through n are all identical.
[0018] The original micro-motion signal time-frequency diagram is processed by the first sub-network to obtain the first mask;
[0019] The data of the i-th mask and the time and frequency of the original micro-motion signal Figure 1 The (i+1)th mask is obtained by processing through the (i+1)th sub-network; the value of i ranges from 1 to n-1.
[0020] The construction of the first sub-network based on the self-attention network specifically includes the following steps:
[0021] The constructed first sub-network includes a linear layer, a self-attention module, a batch normalization module, a dropout module, and an enhancement module;
[0022] The linear layer, self-attention module, batch normalization module, dropout module, enhancement module, and activation function module are connected in series.
[0023] The original time-frequency plot of the input micro-motion signal is position-encoded through a linear layer to obtain Q, K, and V values. The Q, K, and V values are input to a self-attention module to extract features from the input signal. The signal features extracted by the self-attention module are processed by a batch normalization module and a dropout module to obtain output features. The output features are then enhanced by an enhancement module to obtain the first mask.
[0024] The construction of the second to nth subnetworks based on the self-attention network and residual connection scheme specifically includes the following steps:
[0025] The constructed j-th sub-network includes a residual connection module, a linear layer, a self-attention module, a batch normalization module, a dropout module, and an enhancement module; j takes values from 2 to n.
[0026] The residual connection module, linear layer, self-attention module, batch normalization module, dropout module, enhancement module, and activation function module are connected in series.
[0027] The data information of the (j-1)th mask and the original micro-motion signal time-frequency map are processed by the residual connection module, and then position encoded by the linear layer to obtain the Q value; the original micro-motion signal time-frequency map is also directly position encoded by the linear layer to obtain the K value and V value; the Q value, K value and V value are input to the self-attention module to extract the features of the input signal; the signal features extracted by the self-attention module are processed by the batch normalization module and the dropout module to obtain the output features; the output features are then enhanced by the enhancement module to obtain the j-th mask.
[0028] The self-attention module specifically includes the following steps:
[0029] The self-attention module includes a self-attention layer, a first batch normalization layer, a first dropout layer, a first linear layer, an activation layer, a second linear layer, a second batch normalization layer, and a second dropout layer.
[0030] The self-attention layer is based on a multi-head self-attention mechanism. It uses a formula to process the Q-value, K-value, and V-value received by the self-attention module. The data is processed, and the Attention(Q,K,V) data obtained from the processing is concatenated to obtain the first feature map;
[0031] The first feature map is processed by the first batch normalization layer and the first dropout layer to obtain the first feature. The first feature is then processed sequentially by the first linear layer, the activation layer, the second linear layer, the second batch normalization layer, and the second dropout layer to obtain the intermediate feature. The input length of the first linear layer is the spectral width freq of the time-frequency map, which is consistent with the output length of the multi-head attention module. The output length of the first linear layer is 4*freq, which maps the input to a higher-dimensional linear space. The activation layer uses the ReLU function as the activation function to fit the nonlinear transformation. During this process, the data dimension remains unchanged at 4*freq. The input dimension of the second linear layer is 4*freq, and the output dimension is freq, which maps the output back to the original dimension. The intermediate feature is then processed by the second batch normalization layer and the second dropout layer to obtain the second feature. The second feature and the first feature are then concatenated by residuals to obtain the output feature map. The output feature map is the output of the self-attention module.
[0032] The enhancement module specifically includes the following steps:
[0033] The enhancement module includes a first convolutional layer, a first convolutional activation function layer, a second convolutional layer, a second convolutional activation function layer, a third convolutional layer, a third convolutional activation function layer, a fourth convolutional layer, a fourth convolutional activation function layer, a first deconvolutional layer, a first deconvolutional activation function layer, a second deconvolutional layer, a second deconvolutional activation function layer, a third deconvolutional layer, a third deconvolutional activation function layer, a fourth deconvolutional layer, and a fourth deconvolutional activation function layer;
[0034] The output of the first convolutional layer serves as the input to the first convolutional activation function layer; the output of the first convolutional activation function layer serves as the input to the second convolutional layer; the output of the second convolutional layer serves as the input to the second convolutional activation function layer; the output of the second convolutional activation function layer serves as the input to the third convolutional layer; the output of the third convolutional activation function layer serves as the input to the fourth convolutional layer; the output of the fourth convolutional activation function layer serves as the input to the first deconvolutional layer.
[0035] The output of the first deconvolutional layer serves as the input to the first deconvolutional activation function layer. The outputs of the first and third deconvolutional activation functions layers are concatenated and then used as the input to the second deconvolutional layer. The output of the second deconvolutional layer is then used as the input to the second deconvolutional activation function layer. The outputs of the second and third deconvolutional activation functions layers are then concatenated and then used as the input to the third deconvolutional layer. The output of the third deconvolutional layer is then used as the input to the third deconvolutional activation function layer. The outputs of the third and first deconvolutional activation functions layers are then concatenated and then used as the input to the fourth deconvolutional layer. The output of the fourth deconvolutional activation function layer is the output of the enhancement module.
[0036] The first convolutional layer has a kernel size of 3, a stride of 2, and padding of 1; the activation function for the first convolutional activation function layer is ReLU. The second convolutional layer has a kernel size of 3, a stride of 2, and padding of 1; the activation function for the second convolutional activation function layer is ReLU. The third convolutional layer has a kernel size of 3, a stride of 2, and padding of 1; the activation function for the third convolutional activation function layer is ReLU. The fourth convolutional layer has a kernel size of 3, a stride of 2, and padding of 1; the activation function for the fourth convolutional activation function layer is ReLU.
[0037] The first deconvolutional layer has a kernel size of 3, a stride of 2, and padding of 1; the activation function of the first deconvolutional activation function layer is ReLU. The second deconvolutional layer has a kernel size of 3, a stride of 2, and padding of 1; the activation function of the second deconvolutional activation function layer is ReLU. The third deconvolutional layer has a kernel size of 3, a stride of 2, and padding of 1; the activation function of the third deconvolutional activation function layer is ReLU. The fourth deconvolutional layer has a kernel size of 3, a stride of 2, and padding of 1; the activation function of the fourth deconvolutional activation function layer is Sigmoid.
[0038] The residual connection module specifically includes the following steps:
[0039] The following is used as the processing function for the residual connection module in the j-th subnetwork:
[0040]
[0041] In the formula X j X is the output of the residual connection module in the j-th subnetwork; X is the time-frequency diagram of the original micro-motion signal; M i The i-th mask is the output of the i-th subnetwork; ⊙ is the Hadamard product; M j-1ω is the (j-1)th mask output by the (j-1)th subnetwork; ω(0,1) is Gaussian white noise with a mean of 0 and a variance of 1; α is a weight value used to control the noise intensity.
[0042] The training described in step S3 specifically includes the following steps:
[0043] During training, the following function is used as the loss function for model training:
[0044]
[0045] In the formula, L is the value of the loss function; The predicted value output by the model; y jj For the true values in the training dataset; () ii To represent the loss calculated for the ii-th component; Represents the frequencies of N components; Represents the intensity of N components; The phase consists of N components.
[0046] The post-processing described in step S5 specifically includes the following steps:
[0047] By combining the original micro-motion signal time-frequency plot and the mask, a time-frequency plot containing only the corresponding single-component signal is generated, and the corresponding micro-motion curve is directly estimated; the estimation includes estimation using the Viterbi algorithm;
[0048] or
[0049] The center of the mask with a pixel value of 1 is set, and the position of the corresponding micro-motion curve is obtained by fitting the Fourier series least squares method.
[0050] This invention also provides a system for implementing the micro-motion curve extraction method from the time-frequency graph of the micro-motion signal, including a data acquisition module, a model building module, a model training module, a mask extraction module, and a curve extraction module; the data acquisition module, model building module, model training module, mask extraction module, and curve extraction module are connected in series; the data acquisition module is used to acquire existing time-frequency graphs of micro-motion signals and corresponding micro-motion curve data, construct a training dataset, and upload the data information to the model building module; the model building module is used to construct a preliminary mask extraction model for the time-frequency graph of micro-motion signals based on a self-attention network according to the received data information, and upload the data information to the model training module; the model training module is used to... The received data is used to train a preliminary model for mask extraction of the micro-motion signal time-frequency image using the constructed training dataset. This results in a mask extraction model for the micro-motion signal time-frequency image, and the data is then uploaded to the mask extraction module. The mask extraction module takes the actual acquired micro-motion signal time-frequency image and inputs it into the obtained mask extraction model to extract the mask information of the target micro-motion signal time-frequency image. This data is then uploaded to the curve extraction module. The curve extraction module performs post-processing on the mask information of the obtained target micro-motion signal time-frequency image based on the received data to obtain the corresponding micro-motion curve, thus completing the extraction of the micro-motion curve of the target micro-motion signal time-frequency image.
[0051] The present invention provides a method and system for extracting micro-motion curves from time-frequency graphs of micro-motion signals. By constructing a mask extraction model based on a self-attention network, the micro-motion curves are extracted through the extracted mask. Therefore, the present invention can not only extract micro-motion curves from time-frequency graphs of micro-motion signals, but also has higher reliability and better accuracy. Attached Figure Description
[0052] Figure 1 This is a schematic diagram of the method flow of the present invention.
[0053] Figure 2 This is a schematic diagram of a training dataset instance for the method of the present invention.
[0054] Figure 3 This is a schematic diagram of the functional modules of the system of the present invention. Detailed Implementation
[0055] like Figure 1 The diagram shown is a flowchart of the method of the present invention: The method for extracting the micro-motion curve from the time-frequency diagram of the micro-motion signal disclosed in this invention includes the following steps:
[0056] S1. Obtain the existing time-frequency diagram of the micro-motion signal and the corresponding micro-motion curve data, and construct a training dataset;
[0057] S2. Based on a self-attention network, construct a preliminary model for mask extraction of the time-frequency map of micro-motion signals; including the following steps:
[0058] A preliminary model for extracting micro-motion curves is constructed based on a self-attention network.
[0059] The preliminary model for extracting micro-motion curves includes sub-networks 1 to n, where n is the number of micro-motion curves included in the time-frequency plot of the micro-motion signal.
[0060] Construct the first subnetwork based on the self-attention network;
[0061] Based on self-attention networks and residual connection schemes, subnetworks 2 through n are constructed; the structures of subnetworks 2 through n are all identical.
[0062] The original micro-motion signal time-frequency diagram is processed by the first sub-network to obtain the first mask;
[0063] The data of the i-th mask and the time and frequency of the original micro-motion signal Figure 1 The (i+1)th mask is obtained by processing through the (i+1)th sub-network; the value of i ranges from 1 to n-1.
[0064] In specific implementation, the construction of the first sub-network based on the self-attention network includes the following steps:
[0065] The constructed first sub-network includes a linear layer, a self-attention module, a batch normalization module, a dropout module, and an enhancement module;
[0066] The linear layer, self-attention module, batch normalization module, dropout module, enhancement module, and activation function module are connected in series.
[0067] The original time-frequency plot of the input micro-motion signal is position-encoded through a linear layer to obtain Q, K, and V values. The Q, K, and V values are then input into a self-attention module to extract features from the input signal. The signal features extracted by the self-attention module are processed by a batch normalization module and a dropout module to obtain output features. The output features are then enhanced by an enhancement module to obtain the first mask.
[0068] The construction of the second to nth subnetworks based on the self-attention network and residual connection scheme specifically includes the following steps:
[0069] The constructed j-th sub-network includes a residual connection module, a linear layer, a self-attention module, a batch normalization module, a dropout module, and an enhancement module; j takes values from 2 to n.
[0070] The residual connection module, linear layer, self-attention module, batch normalization module, dropout module, enhancement module, and activation function module are connected in series.
[0071] The data information of the (j-1)th mask and the original micro-motion signal time-frequency map are processed by the residual connection module, and then position encoded by the linear layer to obtain the Q value. The original micro-motion signal time-frequency map is also directly position encoded by the linear layer to obtain the K and V values. The Q, K, and V values are input to the self-attention module to extract the features of the input signal. The signal features extracted by the self-attention module are processed by the batch normalization module and the dropout module to obtain the output features. The output features are then enhanced by the enhancement module to obtain the j-th mask. By using two different calculation schemes for the Q, K, and V values, the network is ensured to maintain access to the unprocessed signal information.
[0072] The self-attention module mentioned above specifically includes the following steps:
[0073] The self-attention module includes a self-attention layer, a first batch normalization layer, a first dropout layer, a first linear layer, an activation layer, a second linear layer, a second batch normalization layer, and a second dropout layer.
[0074] The self-attention layer is based on a multi-head self-attention mechanism (preferably with 2 heads). It processes the Q-value, K-value, and V-value received by the self-attention module using the formula... The data is processed, and the Attention(Q,K,V) data obtained from the processing is concatenated to obtain the first feature map;
[0075] The first feature map is processed by the first batch normalization layer and the first dropout layer to obtain the first feature. The first feature is then processed sequentially by the first linear layer, the activation layer, the second linear layer, the second batch normalization layer, and the second dropout layer to obtain the intermediate feature. The input length of the intermediate feature map is the spectral width freq of the time-frequency map, which is consistent with the output length of the multi-head attention module. The output length of the first linear layer is 4*freq, mapping the input to a higher-dimensional linear space. The activation layer uses the ReLU function as the activation function to fit the nonlinear transformation, during which the data dimension remains unchanged at 4*freq. The input dimension of the second linear layer is 4*freq, and the output dimension is freq, mapping the output back to the original dimension. The activation layer also uses the ReLU function as the activation function. The intermediate feature map is then processed by the second batch normalization layer and the second dropout layer to obtain the second feature. The second feature map and the first feature map are then concatenated via residuals to obtain the output feature map, which is the output of the self-attention module.
[0076] The enhancement module specifically includes the following steps:
[0077] The enhancement module includes a first convolutional layer, a first convolutional activation function layer, a second convolutional layer, a second convolutional activation function layer, a third convolutional layer, a third convolutional activation function layer, a fourth convolutional layer, a fourth convolutional activation function layer, a first deconvolutional layer, a first deconvolutional activation function layer, a second deconvolutional layer, a second deconvolutional activation function layer, a third deconvolutional layer, a third deconvolutional activation function layer, a fourth deconvolutional layer, and a fourth deconvolutional activation function layer;
[0078] The output of the first convolutional layer serves as the input to the first convolutional activation function layer; the output of the first convolutional activation function layer serves as the input to the second convolutional layer; the output of the second convolutional layer serves as the input to the second convolutional activation function layer; the output of the second convolutional activation function layer serves as the input to the third convolutional layer; the output of the third convolutional activation function layer serves as the input to the fourth convolutional layer; the output of the fourth convolutional activation function layer serves as the input to the first deconvolutional layer.
[0079] The output of the first deconvolutional layer serves as the input to the first deconvolutional activation function layer. The outputs of the first and third deconvolutional activation functions layers are concatenated and then used as the input to the second deconvolutional layer. The output of the second deconvolutional layer is then used as the input to the second deconvolutional activation function layer. The outputs of the second and third deconvolutional activation functions layers are then concatenated and then used as the input to the third deconvolutional layer. The output of the third deconvolutional layer is then used as the input to the third deconvolutional activation function layer. The outputs of the third and first deconvolutional activation functions layers are then concatenated and then used as the input to the fourth deconvolutional layer. The output of the fourth deconvolutional activation function layer is the output of the enhancement module.
[0080] The first convolutional layer has a kernel size of 3, a stride of 2, and padding of 1; the activation function for the first convolutional activation function layer is ReLU. The second convolutional layer has a kernel size of 3, a stride of 2, and padding of 1; the activation function for the second convolutional activation function layer is ReLU. The third convolutional layer has a kernel size of 3, a stride of 2, and padding of 1; the activation function for the third convolutional activation function layer is ReLU. The fourth convolutional layer has a kernel size of 3, a stride of 2, and padding of 1; the activation function for the fourth convolutional activation function layer is ReLU.
[0081] The first deconvolutional layer has a kernel size of 3, a stride of 2, and padding of 1; the activation function of the first deconvolutional activation function layer is ReLU. The second deconvolutional layer has a kernel size of 3, a stride of 2, and padding of 1; the activation function of the second deconvolutional activation function layer is ReLU. The third deconvolutional layer has a kernel size of 3, a stride of 2, and padding of 1; the activation function of the third deconvolutional activation function layer is ReLU. The fourth deconvolutional layer has a kernel size of 3, a stride of 2, and padding of 1; the activation function of the fourth deconvolutional activation function layer is Sigmoid. The fourth deconvolutional activation function layer outputs a probability distribution map to predict the location of the target signal.
[0082] The residual connection module specifically includes the following steps:
[0083] When the signal is removed from the spectrogram using the output of the previous step, obvious “holes” or “reduced regions” are produced in the image; the network identifies signal boundaries based on changes in the image gradient; such “holes” hinder the convergence of the network, especially when the signal itself contains multiple components; to alleviate this problem, the present invention introduces a residual connection mechanism.
[0084] The following is used as the processing function for the residual connection module in the j-th subnetwork:
[0085]
[0086] In the formula X j X is the output of the residual connection module in the j-th subnetwork; X is the time-frequency diagram of the original micro-motion signal; M i The i-th mask is the output of the i-th subnetwork; ⊙ is the Hadamard product; M j-1 ω(0,1) is the (j-1)th mask output by the (j-1)th subnetwork; ω(0,1) is Gaussian white noise with a mean of 0 and a variance of 1; α is a weight value used to control the noise intensity, preferably 0.3.
[0087] By introducing low-intensity noise into the eliminated regions, the feature maps are made smoother, while gradient information is accumulated, which significantly improves the convergence speed of the model.
[0088] S3. Using the training dataset constructed in step S1, train the preliminary model for mask extraction of the time-frequency image of the micro-motion signal constructed in step S2 to obtain the mask extraction model of the time-frequency image of the micro-motion signal; the training specifically includes the following steps:
[0089] To train the model, the modulation information of the micro-motion signal needs to be considered during data annotation, determining that the components are modulated from low to high in terms of modulation frequency, amplitude, and initial phase. This rule stems from the fact that the order of signal component output is required for calculating the loss function during network training. Therefore, the following function is used as the loss function for model training:
[0090]
[0091] In the formula, L is the value of the loss function; The predicted value output by the model; y jj For the true values in the training dataset; () ii To represent the loss calculated for the ii-th component; Represents the frequencies of N components; Represents the intensity of N components; The phase consists of N components.
[0092] Figure 2 An example training dataset is provided, following the criteria described earlier; it shows a sequence from left to right, starting with the time-frequency distribution (TFD) of the input, followed by the output components in their respective order; in this sequence, blank sections are considered as termination symbols in the output, similar to the <eos> (sentence end) marker used in machine translation tasks; when the model predominantly predicts zeros in its output, it indicates that no more signal components need to be extracted, and it has reached the termination condition; at this node, the model stops further loss calculations and avoids generating additional masked outputs; this design effectively ends signal processing after all important signal components have been identified and extracted, thus preventing unnecessary consumption of computational resources on signal regions that do not contain relevant information;
[0093] S4. Input the actual acquired micro-motion signal time-frequency map into the mask extraction model of the micro-motion signal time-frequency map obtained in step S3, and extract the mask information of the target micro-motion signal time-frequency map;
[0094] S5. Post-process the mask information of the target micro-motion signal time-frequency diagram obtained in step S4 to obtain the corresponding micro-motion curve, and complete the extraction of the micro-motion curve of the target micro-motion signal time-frequency diagram.
[0095] In practice, the post-processing includes the following steps:
[0096] By combining the original micro-motion signal time-frequency plot and the mask, a time-frequency plot containing only the corresponding single-component signal is generated, and the corresponding micro-motion curve is directly estimated; the estimation includes estimation using the Viterbi algorithm;
[0097] or
[0098] The center of the mask with a pixel value of 1 is set, and the position of the corresponding micro-motion curve is obtained by fitting the Fourier series least squares method.
[0099] like Figure 3 The diagram shows the functional modules of the system of the present invention: The system disclosed in this invention for extracting micro-motion curves from the time-frequency graph of a micro-motion signal includes a data acquisition module, a model building module, a model training module, a mask extraction module, and a curve extraction module; these modules are connected in series. The data acquisition module acquires existing time-frequency graphs of micro-motion signals and corresponding micro-motion curve data, constructs a training dataset, and uploads the data information to the model building module. The model building module constructs a preliminary mask extraction model for the time-frequency graph of a micro-motion signal based on a self-attention network, according to the received data information, and uploads the data information to the model training module. The training module trains the preliminary model for mask extraction of the micro-motion signal time-frequency image using the constructed training dataset based on the received data, obtaining the mask extraction model and uploading the data to the mask extraction module. The mask extraction module inputs the actual acquired micro-motion signal time-frequency image into the obtained mask extraction model based on the received data, extracts the mask information of the target micro-motion signal time-frequency image, and uploads the data to the curve extraction module. The curve extraction module performs post-processing on the obtained mask information of the target micro-motion signal time-frequency image based on the received data to obtain the corresponding micro-motion curve, thus completing the micro-motion curve extraction of the target micro-motion signal time-frequency image.
[0100] This invention directly separates the time-frequency distribution of multi-component micro-motion signals into masks corresponding to the positions of individual signal components. Therefore, compared with traditional methods, this invention eliminates the need for manual intervention to determine the number of micro-motion components and perform curve matching and correlation, thus greatly reducing the computational complexity and time cost caused by manual intervention. Furthermore, since this invention is a data-driven separation method, it still possesses a certain degree of separability even for signals with extremely high noise levels, provided the dataset is sufficiently complete, exhibiting higher robustness compared to traditional methods. Finally, because the mask can be combined with the time-frequency distribution to generate the time-frequency distribution corresponding to a single component signal, this invention can also be used to assist in improving the performance of traditional methods.
Claims
1. A method for extracting micro-motion curves from the time-frequency diagram of a micro-motion signal, comprising the following steps: S1. Obtain the existing time-frequency diagrams of micro-motion signals and the corresponding micro-motion curve data, and construct a training dataset; S2. Based on a self-attention network, construct a preliminary model for mask extraction of the time-frequency map of micro-motion signals; including the following steps: A preliminary model for extracting micro-motion curves is constructed based on a self-attention network. The preliminary model for extracting micro-motion curves includes sub-networks 1 to n, where n is the number of micro-motion curves included in the time-frequency plot of the micro-motion signal. Construct the first subnetwork based on the self-attention network; Based on self-attention networks and residual connection schemes, subnetworks 2 through n are constructed; the structures of subnetworks 2 through n are all identical. The original micro-motion signal time-frequency diagram is processed by the first sub-network to obtain the first mask; The data of the i-th mask and the original micro-motion signal time-frequency diagram are processed together through the (i+1)-th sub-network to obtain the (i+1)-th mask; i takes the value from 1 to n-1. S3. Using the training dataset constructed in step S1, train the preliminary model for mask extraction of the time-frequency map of the micro-motion signal constructed in step S2 to obtain the mask extraction model of the time-frequency map of the micro-motion signal. S4. Input the actual acquired micro-motion signal time-frequency map into the mask extraction model of the micro-motion signal time-frequency map obtained in step S3, and extract the mask information of the target micro-motion signal time-frequency map; S5. Post-process the mask information of the target micro-motion signal time-frequency diagram obtained in step S4 to obtain the corresponding micro-motion curve, and complete the extraction of the micro-motion curve of the target micro-motion signal time-frequency diagram. The post-processing specifically includes the following steps: By combining the original micro-motion signal time-frequency plot and the mask, a time-frequency plot containing only the corresponding single-component signal is generated, and the corresponding micro-motion curve is directly estimated; the estimation includes estimation using the Viterbi algorithm; or The center of the mask with a pixel value of 1 is set, and the position of the corresponding micro-motion curve is obtained by fitting the Fourier series least squares method.
2. The method for extracting micro-motion curves from the time-frequency diagram of a micro-motion signal according to claim 1, characterized in that... The construction of the first sub-network based on the self-attention network specifically includes the following steps: The constructed first sub-network includes a linear layer, a self-attention module, a batch normalization module, a dropout module, and an enhancement module; The linear layer, self-attention module, batch normalization module, dropout module, enhancement module, and activation function module are connected in series. The original time-frequency plot of the input micro-motion signal is position-encoded through a linear layer to obtain Q, K, and V values. The Q, K, and V values are then input into a self-attention module to extract features from the input signal. The signal features extracted by the self-attention module are processed by a batch normalization module and a dropout module to obtain the output features. The output features are then enhanced by the enhancement module to obtain the first mask.
3. The method for extracting micro-motion curves from the time-frequency diagram of a micro-motion signal according to claim 2, characterized in that... The construction of the second to nth subnetworks based on the self-attention network and residual connection scheme specifically includes the following steps: The constructed j-th sub-network includes a residual connection module, a linear layer, a self-attention module, a batch normalization module, a dropout module, and an enhancement module; j takes values from 2 to n. The residual connection module, linear layer, self-attention module, batch normalization module, dropout module, enhancement module, and activation function module are connected in series. The data information of the j-1 mask and the original micro-motion signal time-frequency image are processed by the residual connection module, and then position encoded by the linear layer to obtain the Q value; the original micro-motion signal time-frequency image is also directly position encoded by the linear layer to obtain the K value and V value; the Q value, K value and V value are input to the self-attention module to extract the features of the input signal; the signal features extracted by the self-attention module are processed by the batch normalization module and the dropout module to obtain the output features; The output features are then enhanced by the enhancement module to obtain the j-th mask.
4. The method for extracting micro-motion curves from the time-frequency diagram of a micro-motion signal according to claim 3, characterized in that... The self-attention module specifically includes the following steps: The self-attention module includes a self-attention layer, a first batch normalization layer, a first dropout layer, a first linear layer, an activation layer, a second linear layer, a second batch normalization layer, and a second dropout layer. The self-attention layer is based on a multi-head self-attention mechanism. It processes the Q-value, K-value, and V-value received by the self-attention module using the following formula: Process the data and output the processed data. The images are then stitched together to obtain the first feature map. The first feature map is processed by the first batch of normalization layers and the first dropout layer to obtain the first feature; The first feature is then processed sequentially through a first linear layer, an activation layer, a second linear layer, a second batch normalization layer, and a second dropout layer to obtain intermediate features; wherein, the input length of the first linear layer is the spectral width of the time-frequency plot. The output length of the first linear layer is The first linear layer maps the input to a higher-dimensional linear space; the activation layer uses the ReLU function as the activation function to fit the nonlinear transformation; the input dimension of the second linear layer is... The output dimension is This maps the output back to the original dimension; The intermediate features are then processed by a second batch normalization layer and a second dropout layer to obtain the second features; the second features and the first features are then connected by residuals to obtain the output feature map; the output feature map is the output of the self-attention module.
5. The method for extracting micro-motion curves from the time-frequency diagram of a micro-motion signal according to claim 4, characterized in that... The enhancement module specifically includes the following steps: The enhancement module includes a first convolutional layer, a first convolutional activation function layer, a second convolutional layer, a second convolutional activation function layer, a third convolutional layer, a third convolutional activation function layer, a fourth convolutional layer, a fourth convolutional activation function layer, a first deconvolutional layer, a first deconvolutional activation function layer, a second deconvolutional layer, a second deconvolutional activation function layer, a third deconvolutional layer, a third deconvolutional activation function layer, a fourth deconvolutional layer, and a fourth deconvolutional activation function layer; The output of the first convolutional layer serves as the input to the first convolutional activation function layer; the output of the first convolutional activation function layer serves as the input to the second convolutional layer; the output of the second convolutional layer serves as the input to the second convolutional activation function layer; the output of the second convolutional activation function layer serves as the input to the third convolutional layer; the output of the third convolutional activation function layer serves as the input to the fourth convolutional layer; the output of the fourth convolutional activation function layer serves as the input to the first deconvolutional layer. The output of the first deconvolutional layer is used as the input of the first deconvolutional activation function layer. The outputs of the first deconvolutional activation function layer and the third convolutional activation function layer are concatenated and then used as the input of the second deconvolutional layer. The output of the second deconvolutional layer is used as the input of the second deconvolutional activation function layer; The output of the second deconvolutional activation function layer and the output of the second convolutional activation function layer are concatenated and then used as the input of the third deconvolutional layer. The output of the third deconvolutional layer is used as the input of the third deconvolutional activation function layer. The output of the third deconvolutional activation function layer and the output of the first convolutional activation function layer are concatenated and then used as the input of the fourth deconvolutional layer. The output of the fourth deconvolutional layer serves as the input to the fourth deconvolutional activation function layer; the output of the fourth deconvolutional activation function layer serves as the output of the enhancement module. The first convolutional layer has a kernel size of 3, a stride of 2, and padding of 1; the activation function for the first convolutional activation function layer is ReLU. The second convolutional layer has a kernel size of 3, a stride of 2, and padding of 1; the activation function for the second convolutional activation function layer is ReLU. The third convolutional layer has a kernel size of 3, a stride of 2, and padding of 1; the activation function for the third convolutional activation function layer is ReLU. The fourth convolutional layer has a kernel size of 3, a stride of 2, and padding of 1; the activation function for the fourth convolutional activation function layer is ReLU. The first deconvolutional layer has a kernel size of 3, a stride of 2, and padding of 1; the activation function of the first deconvolutional activation function layer is ReLU. The second deconvolutional layer has a kernel size of 3, a stride of 2, and padding of 1; the activation function of the second deconvolutional activation function layer is ReLU. The third deconvolutional layer has a kernel size of 3, a stride of 2, and padding of 1; the activation function of the third deconvolutional activation function layer is ReLU. The fourth deconvolutional layer has a kernel size of 3, a stride of 2, and padding of 1; the activation function of the fourth deconvolutional activation function layer is Sigmoid.
6. The method for extracting micro-motion curves from the time-frequency diagram of a micro-motion signal according to claim 5, characterized in that... The residual connection module specifically includes the following steps: The following is used as the processing function for the residual connection module in the j-th subnetwork: In the formula This is the output of the residual connection module in the j-th subnetwork; This is the time-frequency diagram of the original micro-motion signal; The i-th mask output by the i-th sub-network; For Hadamah accumulation; The (j-1)th mask output by the (j-1)th subnetwork; It is Gaussian white noise with a mean of 0 and a variance of 1; These are the weight values used to control noise intensity.
7. The method for extracting micro-motion curves from the time-frequency diagram of a micro-motion signal according to claim 6, characterized in that... The training described in step S3 specifically includes the following steps: During training, the following function is used as the loss function for model training: In the formula The value of the loss function; The predicted value output by the model; The true values in the training dataset; To indicate the first The loss is calculated from the components; express The frequency of each component; express The intensity of each component; for The phase of each component.
8. A system for implementing the method for extracting micro-motion curves from the time-frequency diagram of a micro-motion signal as described in any one of claims 1 to 7, characterized in that... It includes a data acquisition module, a model building module, a model training module, a mask extraction module, and a curve extraction module; these modules are connected in series. The data acquisition module acquires existing micro-motion signal time-frequency diagrams and corresponding micro-motion curve data, builds a training dataset, and uploads the data information to the model building module. The model building module, based on the received data information and a self-attention network, builds a preliminary mask extraction model for the micro-motion signal time-frequency diagram and uploads the data information to the model training module. The model training module is used to train the preliminary model for mask extraction of the micro-motion signal time-frequency image based on the received data information and the constructed training dataset, thereby obtaining the mask extraction model for the micro-motion signal time-frequency image, and uploading the data information to the mask extraction module; the mask extraction module is used to input the actually acquired micro-motion signal time-frequency image into the obtained mask extraction model for the micro-motion signal time-frequency image based on the received data information, extract the mask information of the target micro-motion signal time-frequency image, and upload the data information to the curve extraction module; The curve extraction module is used to post-process the mask information of the target micro-motion signal time-frequency diagram based on the received data information to obtain the corresponding micro-motion curve, thus completing the extraction of the micro-motion curve from the target micro-motion signal time-frequency diagram.