Radar signal state classification method and device based on time sequence labeling and storage medium

Abstract

The invention discloses a radar signal state classification method and device based on time sequence labeling and a storage medium, and the method comprises the steps: carrying out the column division of radar signal data according to the arrival time features TOA of two data points before and after scanning and tracking a radar signal TWS; extracting trend features and seasonal features of the radar signals through a Transform module, and generating a target sequence; performing characterization learning on the target sequence by using a double-tower encoder, and fusing characterization results through a gating unit; marking each data point in the time sequence, and judging the state of each data point; and a regularization decoder is used to predict the consistency of adjacent labels in the time sequence, the incoherent conversion of the labels in the time sequence is judged through a regularization item and a total loss function, and the radar signals are classified.

Description

Technical Field

[0001] The present invention relates to the technical field of time series and signal classification, and in particular to a radar signal state classification method, device and storage medium based on time series annotation. Background Art

[0002] With the continuous development of radar technology, multi-function radar systems, such as track-while-scan (TWS) radars, have been widely deployed on the battlefield. In order to maintain an advantage in radar confrontation, accurately predicting the operating state of the target radar (e.g., tracking, searching, etc.) has become an urgent need. Since the signal patterns emitted by radars in different states are different, the radar field has gradually derived research on radar signal state classification. This study aims to classify the received radar signal by marking the corresponding state for each signal point in the sub-segment. Considering the difficulty of manually understanding radar signals, radar signal state classification methods based on artificial intelligence models have emerged. However, due to the particularity of the radar confrontation field, the real data of radar signals are scarce due to their confidential nature, so model training requires the use of a large amount of simulation data, and simulation data does not have the non-balanced characteristics of real data. Therefore, the existing data sets have the problems of small amount of real data and low quality of simulation data. On the other hand, the duration of each state in the radar signal is not uniform, and it is impossible to use a fixed-length window segment for sequence classification. Therefore, how to perform length adaptive analysis on the sequence is another problem that needs to be solved.

[0003] In summary, the radar signal state classification task has the following difficulties:

[0004] 1. The amount of radar signal data is small and the simulation sequence quality is low;

[0005] 2. Radar signals have non-equilibrium characteristics, and it is difficult for the model to adaptively segment them;

[0006] The existing technology is difficult to solve the above problems well, so the state classification of radar signals is not easy. Summary of the invention

[0007] The purpose of the present invention is to provide a radar signal state classification method, device and storage medium based on time series annotation to solve the data set quality problem and sequence length adaptive analysis problem, aiming to improve the classification effect of the model in real scenarios.

[0008] To achieve the above object, the present invention provides the following technical solutions:

[0009] A radar signal state classification method based on time series annotation, comprising:

[0010] Data augmentation module;

[0011] According to the arrival time characteristics TOA of the two data points before and after the scanning and tracking radar signal TWS, the radar signal data is divided into columns;

[0012] The trend and seasonal characteristics of radar signals are extracted through the Transformer module to generate the target sequence;

[0013] Timing annotation module;

[0014] The dual-tower encoder is used to learn the representation of the target sequence, and the representation results are fused through the gating unit;

[0015] Label each data point in the time series and determine the status of each data point;

[0016] The regularized decoder is used to predict the consistency of adjacent labels in the time series. The incoherent transitions of labels in the time series are judged by the regularization term and the total loss function to classify the radar signals.

[0017] Preferably, the power amplification feature PA of the radar signal is extracted using the following expression, and the difference between two data points in the arrival time feature TOA is calculated based on the power amplification feature PA;

[0018] PA(f,θ,dis)=G(f,θ)+A(dis)+N(0,3)+290;

[0019] in,

[0020]

[0021] A(dis)=-10×log 10 (dis 2 );

[0022] Where N(0,3) is Gaussian noise with a mean of 0 and a standard deviation of 3; θ is the effective angle calculated from the elevation angle and azimuth angle; G(f,θ) is the antenna gain; A(dis) is the gain loss caused by the distance; dis is the distance; ang el is the pitch angle; ang az is the azimuth; f is the radio frequency; j is the imaginary unit; λ is the signal wavelength; d is the distance between adjacent array elements; θ B is the pointing angle of the beam.

[0023] Preferably, the Transformer module includes an encoder and a decoder;

[0024] In the encoder, a diffusion model is used to jointly model and encode the target sequence and the conditional sequence through a cross-attention mechanism;

[0025] The decoder adopts a multi-layer structure, each layer of which contains a Transformer block, a feed-forward neural network and a set of feature extraction layers;

[0026] The feature extraction layer includes a trend extractor and a seasonality extractor.

[0027] Preferably, the joint modeling coding comprises the following steps:

[0028] First, perform linear transformation on the two sets of data points in the radar signal, and assign query vector Query, key vector Key, and value vector Value to the two sets of data points. The formula is as follows:

[0029] Q=XW Q ,K=CW K ,V=CW V ;

[0030]

[0031] Where W Q ,W K ,W V are the weight matrices of Q, K, and V respectively; X is the target sequence data; n is the number of data points; d x is the feature dimension of each data point; C is the conditional data; d c is the feature dimension of each conditional data point;

[0032] Secondly, the attention score is calculated using the query vector Query and the key vector Key. The formula is as follows:

[0033]

[0034] Finally, the final joint representation output w, w = A·V, is obtained through the Value value and the attention score, which is used as the joint encoding of the target sequence and the conditional sequence for subsequent feature extraction.

[0035] Preferably, the calculation expression of the trend extractor is:

[0036]

[0037] In the formula, is the average value of the output of the i-th decoder block; w i,t is the input of the trend feature extraction module by the i-th decoder block at time step t, which contains the encoding information of the target sequence and the conditional sequence through the cross attention mechanism; Linear is a linear transformation; C is a slowly changing polynomial space matrix; p is the degree of the polynomial, that is, the degree of the highest order term;

[0038] The calculation expression of the seasonal extractor is:

[0039]

[0040] In the formula, is the amplitude A of the seasonal component of the kth frequency of the ith decoder block at time step t; is the Fourier transform function; k is the frequency; is the phase of the seasonal component of the kth frequency of the i-th decoder block at time step t; φ is the function of the phase angle of the calculated frequency component; TopK is the first K selected; is the amplitude A of the first K largest seasonal components of the kth frequency of the ith decoder block at time step t; τ is the length of one cycle of the time series; S i,t is the seasonal pattern reconstructed by the i-th decoder block at time step t; is the amplitude of the first k largest seasonal components of the ith decoder block at time step t; is the positive frequency component, which indicates the frequency of a specific seasonal pattern in the seasonal analysis of the time series; c is the normalized frequency, which is used to adjust the frequency component to match the actual periodicity of the time series data; is the phase of the first k largest seasonal components of the ith decoder block at time step t; is the negative frequency component; in the Fourier transform of a real signal, the negative frequency component is the complex conjugate of the positive frequency component, and together they constitute the complete spectrum of the signal; is the phase of the negative frequency component.

[0041] Preferably, the dual-tower encoder includes a time-step feature encoder and a multivariate feature encoder; the dual-tower encoder performs characterization learning on the target sequence including the following steps:

[0042] Time step feature encoder: Positionally encode the time series, input the positional encoding and the time series data obtained by nonlinear transformation encoding into the time step feature encoder, use self-attention and mask mechanisms to focus on each point on a single channel, and calculate the pairwise attention weights between all time steps through the QKV mode (query vector, key vector, value vector);

[0043] Multivariate feature encoder: It captures the correlation between multiple variables based on the attention mechanism on all channels of the current time step through masks;

[0044] A fully connected layer with nonlinear activation is used to process in the order of time steps or sequences, and the time step feature output is mapped into a step-by-step manner, and each feature channel is processed independently; the multivariate feature output is mapped into a channel-wise manner; the time step and multivariate features are mapped into a vector h through concatenation and linear projection, and each dimension in the vector h is mapped into the weight of each tower through Softmax, and finally the final representation is obtained by the weight and the output of each tower.

[0045] Preferably, the regularization term is the sum of the conversion costs between adjacent labels:

[0046]

[0047] in,

[0048] ρ(y t ,y t+1 )=1-δ(y t ,y t+1 );

[0049] In the formula, ρ(y t ,y t+1 ) is the cost function, which is used to quantify the change from state y t Transition to state y t+1 The inconsistency of δ is the Kronecker function if and only if y t =y t+1 Its value is 1 when it is, otherwise it is 0;

[0050] The total loss function is the weighted sum of the prediction loss and the regularization term:

[0051]

[0052] In the formula, For model-based prediction and the prediction loss of the true label Y; λ is the regularization weight, which is used to control the influence of the regularization term on the total loss.

[0053] A computer device comprises a memory and a processor, wherein the memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the radar signal state classification method based on timing annotation.

[0054] A computer storage medium is used to store program data. When the program data is executed by a processor, it is used to implement the radar signal state classification method based on time series annotation.

[0055] Compared with the prior art, the present invention has the following beneficial effects:

[0056] The present invention performs data enhancement on radar signal simulation data through a method based on a diffusion model and a Transformer module, wherein the columnar features of the radar signal are divided into columns using a sequence method to achieve a more accurate generation effect of the target sequence; subsequently, by introducing the sequence labeling task into the field of time series classification, the state classification task of the non-equilibrium radar signal is completed with a higher accuracy, and the model performance is improved through effective regularization. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is a schematic diagram of the overall workflow of the present invention;

[0058] Figure 2 This is a schematic diagram of a radar signal sequence according to an embodiment of the present invention;

[0059] Figure 3 This is a schematic diagram of radar signal training Batch comparison according to an embodiment of the present invention;

[0060] Figure 4 A schematic diagram of the structure of a decoder according to an embodiment of the present invention;

[0061] Figure 5 Schematic diagram of regularization of time series annotation based on Transformer according to an embodiment of the present invention;

[0062] Figure 6 It is a schematic diagram of an example of implementing a task of the present invention. DETAILED DESCRIPTION

[0063] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0064] Embodiment 1:

[0065] See also Figure 1-Figure 5 ,A radar signal state classification method based on time series annotation, comprising the following steps;

[0066] One, such as Figure 1 As shown in the figure, based on the diffusion model, the data enhancement module is used to perform style transfer on the signal simulation sequence to generate a target sequence that is closer to the real data in distribution, so as to improve the quality and quantity of the dataset.

[0067] like Figure 2As shown in the figure, the scanning and tracking radar signal TWS can be regarded as a multivariate time series, including two main features: power amplification PA and arrival time TOA. Among them, the arrival time feature TOA has an obvious step-like feature. Therefore, the sequence drawn with the arrival time feature TOA as the horizontal axis and the power amplification feature PA as the vertical axis has a non-balanced feature and presents a multi-column feature. Within each column, the power amplification feature PA shows an envelope trend that first rises and then falls, while the arrival time feature TOA shows a gradually rising step trend. Due to the large difference in values ​​between each column, the slow increase trend in the column and the surge trend between columns cannot be well learned by the model at the same time, so a column splitting strategy is adopted, such as Figure 3 As shown, the model learns the changing characteristics of the time series in each column to better fit the changing trend and periodicity of the sequence.

[0068] The power amplification characteristic PA involves multiple physical parameters such as radio frequency, pitch angle, azimuth angle and distance. The power amplification characteristic PA reflects the state change trend and periodicity of the radar signal. The calculation expression is as follows:

[0069] PA(f,θ,dis)=G(f,θ)+A(dis)+N(0,3)+290;

[0070] in,

[0071]

[0072] A(dis)=-10×log 10 (dis 2 );

[0073] Where N(0,3) is Gaussian noise with a mean of 0 and a standard deviation of 3; θ is the effective angle calculated from the elevation angle and azimuth angle; G(f,θ) is the antenna gain; A(dis) is the gain loss caused by the distance; dis is the distance; ang el is the pitch angle; ang az is the azimuth; f is the radio frequency; j is the imaginary unit; λ is the signal wavelength; d is the distance between adjacent array elements; θ B is the pointing angle of the beam.

[0074] Specifically, a mutation threshold of the TOA value is established based on the signal value. When the difference between the two data points in the TOA dimension of the arrival time feature is greater than the threshold, different columns are divided between the two data points. After the division is completed, the batch_size of the diffusion model training is designed according to the amount of data in each column, so that the divided single signal columns are in the same batch, allowing the model to focus on the data feature learning and simulation generation within a single column, so as to better adapt to the problem of sudden increase of data between non-equilibrium sequences.

[0075] Second, the time series features are decoupled and learned through the encoder-decoder based Transformer module, enhancing the model's ability to capture global correlations, time series trends, and periodic features.

[0076] In the encoder, the target sequence and the conditional sequence are jointly modeled and encoded through a cross-attention mechanism, and a diffusion model is adopted.

[0077] Specifically,

[0078] First, based on the attention mechanism, the two sets of data points in the radar signal are linearly transformed, and the two sets of data points are assigned query vector Query, key vector Key, and value vector Value. The formula is as follows:

[0079] Q=XW Q ,K=CW K ,V=CW V ;

[0080]

[0081] Where W Q ,W K ,W V are the weight matrices of Q, K, and V respectively; X is the target sequence data; n is the number of data points; d x is the feature dimension of each data point; C is the conditional data; d c is the feature dimension for each conditional data point.

[0082] Secondly, the attention score is calculated using the query vector Query and the key vector Key. The formula is as follows:

[0083]

[0084] These attention scores determine how each target sequence data point implicitly focuses on constraints such as the smoothness of state changes and the correlation between different state segments.

[0085] Finally, the final joint representation output w, w = A·V, is obtained through the Value value and the attention score, which is used as the joint encoding of the target sequence and the conditional sequence for subsequent feature extraction.

[0086] like Figure 4 As shown, the decoder adopts a multi-layer structure, each layer includes a Transformer block, a feedforward neural network and a set of feature extraction layers; the feature extraction layer includes a trend extractor and a seasonality extractor.

[0087] After obtaining the joint representation output, the trend feature is used to capture the overall trend of the non-stationary sequence in a smoothly changing state. The trend feature is modeled using the multinomial regressor method. The formula is as follows:

[0088]

[0089] In the formula, is the average value of the output of the i-th decoder block; w i,t is the input of the i-th decoder block to the trend feature extraction module at time step t, which contains the encoding information of the target sequence and the conditional sequence through the cross attention mechanism; Linear is a linear transformation; C is a slowly changing polynomial space matrix; p is the degree of the polynomial, that is, the degree of the highest order term.

[0090] Seasonality includes periodic features in the sequence and non-periodic error features. The seasonal features of the time series in the Fourier synthesis layer are captured using the Fourier series. The Fourier series can transform the time series data from the time domain to the frequency domain to better analyze and understand the periodic patterns in the data, and extract frequency components for synthesizing or reconstructing time series with specific seasonal features. The calculation expression of the seasonal extractor is:

[0091]

[0092] In the formula, is the amplitude A of the seasonal component of the kth frequency of the i-th decoder block at time step t; F is the Fourier transform function; k is the frequency; is the phase of the seasonal component of the kth frequency of the i-th decoder block at time step t; φ is the function of the phase angle of the calculated frequency component; TopK is the first K selected; is the amplitude A of the first K largest seasonal components of the kth frequency of the ith decoder block at time step t; τ is the length of one cycle of the time series; S i,t is the seasonal pattern reconstructed by the i-th decoder block at time step t; is the amplitude of the first k largest seasonal components of the ith decoder block at time step t; is the positive frequency component, which indicates the frequency of a specific seasonal pattern in the seasonal analysis of the time series; c is the normalized frequency, which is used to adjust the frequency component to match the actual periodicity of the time series data; is the phase of the first k largest seasonal components of the ith decoder block at time step t; is the negative frequency component; in the Fourier transform of a real signal, the negative frequency component is the complex conjugate of the positive frequency component, and together they constitute the complete spectrum of the signal; is the phase of the negative frequency component.

[0093] 3. Figure 5 As shown in the figure, the Transformer-based temporal annotation regularization module represents the time series, the dual-tower encoder module is used to fuse the representation results through the gating unit, and the regularization model is used to guide the model annotation training based on the continuity of the temporal segment state.

[0094] The dual-tower encoder module includes a time-step feature encoder and a multivariate feature encoder.

[0095] In the time-step feature encoder, in order to encode the time features, self-attention and mask are used to focus on each point on a single channel, and the pairwise attention weights between all time steps are calculated through QKV. Since the time-step features are affected by the sequence order, the sequence is first position-encoded, and the position encoding and the time series data obtained by nonlinear transformation encoding are input into the encoder together; in the multivariate feature encoder, the encoder captures the correlation between multiple variables based on the attention mechanism on all channels of the current time step through the mask.

[0096] After implementing the dual-tower encoding, in order to merge and encode the features of two dimensions, a gating mechanism is used to learn the weight of each tower. After obtaining the output of each tower, a fully connected layer with nonlinear activation is used to map the time step feature output to step-wise and the multi-feature output to channel-wise. The time step feature and multi-feature are mapped to a vector h through concatenation and linear projection, and each dimension in h is mapped to the weight of each of the two towers through Softmax. Finally, the final representation is obtained by the weight and the output of each tower.

[0097] like Figure 6 As shown in the figure, due to the continuity and smoothness of state transitions in time series, adjacent data points are often more likely to be in the same state category. However, the standard loss function in sequence labeling tasks usually fails to fully consider the characteristics of time continuity and state transition smoothness. In order to effectively model and process this type of characteristics, a regularized time series labeling model is established. The regularized time series labeling model pays special attention to enhancing the consistency of the model when predicting adjacent labels. By penalizing incoherent transitions in the label sequence, the regularized time series labeling model is strengthened to learn the state continuity in the time series, thereby improving the overall performance of the sequence labeling task.

[0098] Specifically, for the sequence X = x1, x2, …, x T , let Y = y1, y2, …, y T represents the output label sequence of sequence labeling, where y tis the state category label of the data point at time step t. This is achieved by introducing an additional regularization term, which is defined as the sum of the conversion costs between adjacent labels, quantifying the conversion costs between consecutive labels in the sequence. In this way, the time series annotation model not only learns to identify the correct state of each independent data point, but is also guided to understand and predict state transitions in a way that is more consistent with the actual sequence dynamics, thereby reducing misjudgments and improving overall annotation consistency without sacrificing response speed. The regularization term is:

[0099]

[0100] in,

[0101] ρ(y t ,y t+1 )=1-δ(y t ,y t+1 );

[0102] In the formula, ρ(y t ,y t+1 ) is the cost function, which is used to quantify the change from state y t Transition to state y t+1 The inconsistency of δ is the Kronecker function if and only if y t =y t+1 Its value is 1 when , otherwise it is 0. This cost function only generates costs when there is a change between labels, thus encouraging adjacent labels to remain consistent.

[0103] During the training process of the time series labeling model, the total loss function is defined as the weighted sum of the prediction loss and the regularization term:

[0104]

[0105] In the formula, For model-based prediction and the prediction loss of the true label Y; λ is the regularization weight, which is used to control the influence of the regularization term on the total loss.

[0106] Regularization strategies not only help time series annotation models capture more subtle state changes, but also play a role in suppressing overfitting during model training, especially when the labeled data is sparse or unbalanced. By properly adjusting the regularization weights, the emphasis on state consistency can be flexibly controlled, making model training more robust and further improving the model's generalization ability for complex time series data.

[0107] Time series analysis in the radar field is essentially the analysis of time series data that reflects the interaction between the target object and the radar sensor. It has great application value in accurately predicting the operating status of the target radar (for example, tracking, searching, etc.). By classifying the status of the radar at different times, it is possible to effectively determine whether the radar is in a normal working environment and accurately identify when it is subject to external interference in order to deploy relevant strategies.

[0108] Embodiment 2:

[0109] The present application also provides a computer storage medium, which is used to store program data. When the program data is executed by a processor, it is used to implement the radar signal state classification method based on time series annotation as described in the above embodiment.

[0110] The present application also provides a computer program product, wherein the computer program product includes a computer program, and the computer program is operable to enable a computer to execute the radar signal state classification method based on time series annotation as described in the embodiment of the present application. The computer program product can be a software installation package.

[0111] The radar signal state classification method based on timing annotation described in the above embodiment of the present application, when implemented in the form of a software functional unit and sold or used as an independent product, can be stored in a device, such as a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or part of the contribution to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor (processor) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program code.

[0112] In a specific embodiment, Table 1 is a radar signal state classification performance evaluation table.

[0113] Table 1 Radar signal status classification performance evaluation table

[0114]

[0115]

[0116] The present invention (MEDIAN) is significantly better than other baseline models in multiple evaluation indicators. In terms of key performance indicators such as accuracy, precision, recall and F1 value, the present invention shows a significant improvement of 41.8% compared with the KNN model, 26.2% compared with the LR model, 36.3% compared with the MLP model, 8.2% compared with the LSTM model, and even 7.0% compared with the advanced TST model. This result not only highlights the superiority of the present invention in the radar signal state classification task, but also confirms that the robustness and generalization ability of the model can be significantly improved by data enhancement and the introduction of regularization terms that describe the consistency of the time series state.

[0117] The above embodiments are merely examples for the purpose of clear explanation, and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived from these are still within the scope of protection of the invention.

Claims

1. A radar signal state classification method based on time series annotation, characterized in that: include: Data augmentation module; According to the arrival time characteristics TOA of the two data points before and after the scanning and tracking radar signal TWS, the radar signal data is divided into columns; The trend and seasonal characteristics of radar signals are extracted through the Transformer module to generate the target sequence; Timing annotation module; The dual-tower encoder is used to learn the representation of the target sequence, and the representation results are fused through the gating unit; Label each data point in the time series and determine the status of each data point; The regularized decoder is used to predict the consistency of adjacent labels in the time series. The incoherent transitions of labels in the time series are judged by the regularization term and the total loss function to classify the radar signals.

2. The radar signal state classification method based on time series annotation according to claim 1 is characterized in that: The power amplification feature PA of the radar signal is extracted using the following expression, and the difference between two data points in the arrival time feature TOA is calculated based on the power amplification feature PA; PA(f,θ,dis)=G(f,θ)+A(dis)+N(0,3)+290; in, A(dis)=-10×log 10 (dis 2 ); Where N(0,3) is Gaussian noise with a mean of 0 and a standard deviation of 3; θ is the effective angle calculated from the elevation angle and azimuth angle; G(f,θ) is the antenna gain; A(dis) is the gain loss caused by the distance; dis is the distance; ang el is the pitch angle; ang az is the azimuth; f is the radio frequency; j is the imaginary unit; λ is the signal wavelength; d is the distance between adjacent array elements; θ B is the pointing angle of the beam.

3. The radar signal state classification method based on time series annotation according to claim 1 is characterized in that: The Transformer module includes an encoder and a decoder; In the encoder, a diffusion model is used to jointly model and encode the target sequence and the conditional sequence through a cross-attention mechanism; The decoder adopts a multi-layer structure, each layer of which contains a Transformer block, a feed-forward neural network and a set of feature extraction layers; The feature extraction layer includes a trend extractor and a seasonality extractor.

4. The radar signal state classification method based on time series annotation according to claim 3 is characterized in that: The joint modeling coding comprises the following steps: First, perform linear transformation on the two sets of data points in the radar signal, and assign query vector Query, key vector Key, and value vector Value to the two sets of data points. The formula is as follows: Q=XW Q ,K=CW K ,V=CW v ; Where W Q ,W K ,W V are the weight matrices of Q, K, and V respectively; X is the target sequence data; n is the number of data points; d x is the feature dimension of each data point; C is the conditional data; d c is the feature dimension of each conditional data point; Secondly, the attention score is calculated using the query vector Query and the key vector Key. The formula is as follows: Finally, the final joint representation output w, w = A·V, is obtained through the Value value and the attention score, which is used as the joint encoding of the target sequence and the conditional sequence for subsequent feature extraction.

5. The radar signal state classification method based on time series annotation according to claim 3 is characterized in that: The calculation expression of the trend extractor is: In the formula, is the average value of the output of the i-th decoder block; w i,t is the input of the trend feature extraction module by the i-th decoder block at time step t, which contains the encoding information of the target sequence and the conditional sequence through the cross attention mechanism; Linear is a linear transformation; C is a slowly changing polynomial space matrix; p is the degree of the polynomial, that is, the degree of the highest order term; The calculation expression of the seasonal extractor is: In the formula, is the amplitude A of the seasonal component of the kth frequency of the ith decoder block at time step t; is the Fourier transform function; k is the frequency; is the phase of the seasonal component of the kth frequency of the i-th decoder block at time step t; φ is the function of the phase angle of the calculated frequency component; TopK is the first K selected; is the amplitude A of the first K largest seasonal components of the kth frequency of the ith decoder block at time step t; τ is the length of one cycle of the time series; S i,t is the seasonal pattern reconstructed by the i-th decoder block at time step t; is the amplitude of the first k largest seasonal components of the ith decoder block at time step t; is the positive frequency component, which indicates the frequency of a specific seasonal pattern in the seasonal analysis of the time series; c is the normalized frequency, which is used to adjust the frequency component to match the actual periodicity of the time series data; is the phase of the first k largest seasonal components of the ith decoder block at time step t; is the negative frequency component; in the Fourier transform of a real signal, the negative frequency component is the complex conjugate of the positive frequency component, and together they constitute the complete spectrum of the signal; is the phase of the negative frequency component.

6. The radar signal state classification method based on time series annotation according to claim 1 is characterized in that: The dual-tower encoder includes a time-step feature encoder and a multivariate feature encoder; The dual-tower encoder learns the representation of the target sequence by the following steps: Time step feature encoder: Position encoding is performed on the time series, and the position encoding and the time series data obtained by nonlinear transformation encoding are input into the time step feature encoder together. Self-attention and mask mechanisms are used to focus on each point on a single channel, and pairwise attention weights between all time steps are calculated. Multivariate feature encoder: It captures the correlation between multiple variables based on the attention mechanism on all channels of the current time step through masks; A fully connected layer with nonlinear activation is used to process in the order of time steps or sequences, and the time step feature output is mapped into a step-by-step manner, and each feature channel is processed independently; the multivariate feature output is mapped into a channel-wise manner; the time step and multivariate features are mapped into a vector h through concatenation and linear projection, and each dimension in the vector h is mapped into the weight of each tower through Softmax, and finally the final representation is obtained by the weight and the output of each tower.

7. The radar signal state classification method based on time series annotation according to claim 1 is characterized in that: The regularization term is the sum of the conversion costs between adjacent labels: in, ρ(and t ,and t+1 )=1-δ(y t ,and t+1 ); In the formula, ρ(y t ,y t+1 ) is the cost function, which is used to quantify the change from state y t Transition to state y t+1 The inconsistency of δ is the Kronecker function if and only if y t =y t+1 Its value is 1 when it is, otherwise it is 0; The total loss function is the weighted sum of the prediction loss and the regularization term: In the formula, For model-based prediction and the prediction loss of the true label Y; λ is the regularization weight, which is used to control the influence of the regularization term on the total loss.

8. A computer device, characterized in that: The device includes a memory and a processor, wherein the memory is coupled to the processor; the memory is used to store program data, and the processor is used to execute the program data to implement the radar signal state classification method based on timing annotation as described in any one of claims 1 to 7.

9. A computer storage medium, characterized in that The computer storage medium is used to store program data, and when the program data is executed by the processor, it is used to implement the radar signal state classification method based on time series marking according to any one of claims 1 to 7.

Images