Radar working mode single-step prediction method based on time sequence fusion Transform

Abstract

The invention provides a radar working mode single-step prediction method based on time sequence fusion Transform, belongs to the technical field of radar signal processing, solves the problems of low prediction precision and poor prediction function robustness in traditional radar working mode prediction, and specifically comprises the following steps: carrying out transform domain analysis on an obtained radar pulse signal to obtain a prediction result; constructing a radar working mode original sample data set; performing multi-scale feature extraction and recognition on the time-frequency graph in the radar working mode original sample data set, and constructing a training data set based on extracted features and corresponding radar working mode labels obtained through recognition; constructing a TFT prediction model and training the TFT prediction model by using the training data set; performing transform domain analysis and multi-scale feature extraction on the received radar signals to generate historical features and tag sequences; and inputting the generated historical features and the tag sequence into the trained TFT prediction model to carry out single-step prediction of the radar working mode, and outputting a prediction result.

Description

Technical Field

[0001] The invention relates to a single-step prediction method for a radar working mode based on a time series fusion Transformer, and belongs to the technical field of radar signal processing. Background Art

[0002] With the rapid development of phased array radar technology, its flexible beam pointing, waveform modulation and multi-task processing capabilities have greatly improved the adaptability and complexity of radar systems, which has brought huge technical challenges to radar reconnaissance and confrontation. Radar working mode prediction is one of the important technologies in the field of electronic warfare. It aims to predict the future working state of the radar in advance by analyzing the intercepted radar pulse signal. The radar working mode prediction algorithm needs to mine the inherent laws of radar behavior based on the timing characteristics of the pulse signal on the basis of identifying the historical working mode, so as to achieve more accurate and reliable working mode prediction.

[0003] Temporal Fusion Transformer (TFT) is an architecture designed for time series prediction tasks. It has the ability to capture time series dependencies and is particularly suitable for complex time series analysis. The TFT prediction model combines four modules: gated residual network, variable selection network, static covariate encoder, and temporal fusion decoder, which can effectively process static and dynamic features. Through staged processing, the most significant and core features are dynamically selected, and global and local temporal dependencies are used to predict the target sequence. Summary of the invention

[0004] In order to solve the problems of low prediction accuracy and poor prediction function robustness in traditional radar working mode prediction, the present invention proposes a radar working mode single-step prediction method based on time series fusion Transformer, and the specific steps include:

[0005] Step 1: Perform transform domain analysis on the acquired radar pulse signal to construct the original sample data set of the radar working mode;

[0006] Step 2: Extract and identify multi-scale features of the time-frequency graph in the original sample data set of radar working modes, and construct a training data set based on the extracted features and the corresponding radar working mode labels obtained by identification;

[0007] Step 3: Build a TFT prediction model and use the training data set to train the TFT prediction model;

[0008] Step 4: Perform transform domain analysis and multi-scale feature extraction on the received radar signal to generate historical features and label sequences;

[0009] Step 5: Input the generated historical features and label sequences into the trained TFT prediction model to perform single-step prediction of the radar working mode and output the prediction results.

[0010] Preferably, the step of constructing the radar working mode original sample data set in step 1 includes:

[0011] Step 1.1: Use the WVD formula to calculate the energy distribution of the acquired radar pulse signal in frequency at each time segment, complete the time-frequency analysis, and obtain a WVD time-frequency image in the form of a two-dimensional matrix, where the radar pulse signal is an IQ signal in the time domain, I is the in-phase component, Q is the orthogonal component, and the IQ signal contains the status information of the radar in different working modes;

[0012] Step 1.2: Adjust the WVD time-frequency image in the form of a two-dimensional matrix to a size of 224*224*3 to complete the WVD time-frequency image standardization process, where the three channels have different frequency bandwidths and time domain features;

[0013] Step 1.3: The WVD time-frequency image is annotated according to the corresponding radar working mode to generate a labeled sample. The standardized WVD time-frequency image is classified according to the radar working model to construct the radar working mode original sample data set;

[0014] The WVD formula is:

[0015]

[0016] In formula (1), W x (t,f) is the WVD time-frequency representation of the signal at time t and frequency f, x * (t) is the complex conjugate of the signal, τ is the delay variable in the time domain, which represents the time difference used in the integration process, e -j2πfτ is the frequency component in Fourier transform.

[0017] Preferably, in step 2, performing multi-scale feature extraction and recognition on the time-frequency graph in the original sample data set of the radar working mode specifically includes:

[0018] The original sample dataset of radar working mode is input into the multi-scale feature extraction architecture based on the BiFormer model to extract multi-scale features. The extracted multi-scale features are input into the classification module composed of layer normalization layer, average pooling layer, tensor flattening layer and classification head to obtain the corresponding radar working mode label.

[0019] Preferably, the multi-scale feature extraction architecture based on the BiFormer model includes a Patch embedding module, a BiFormer module and a Patch merging module;

[0020] The Patch embedding module is used to divide the input time-frequency graph into different sizes;

[0021] The BiFormer module is a core building block based on a two-level routing attention mechanism, used to extract multi-scale features from the input feature map;

[0022] The Patch merging module is used to reduce the size of the feature map processed by the BiFormer module.

[0023] Extracting multi-scale features specifically includes:

[0024] In the initial feature extraction stage, the Patch embedding module divides the input time-frequency map into a set of small image patches, stacks and convolves the time-frequency map along the channel direction, divides it into initial feature patches of size 56×56×64, and processes them through two BiFormer modules;

[0025] In the intermediate feature extraction stage, the feature map size after being processed by the two BiFormer modules is reduced to 28×28×128 and 14×14×256 through the Patch merging module. The feature map with a size of 28×28×128 is processed by two BiFormer modules, and the feature map with a size of 14×14×256 is processed by eight BiFormer modules.

[0026] In the high-level feature extraction stage, the feature map sizes processed by two BiFormer modules and eight BiFormer modules are uniformly reduced to 7×7×512 through the Patch merging module and processed by two BiFormer modules to complete multi-size feature extraction.

[0027] Preferably, the TFT prediction model constructed in step 3 includes a variable selection network, a static covariate encoder, a gated residual network and a temporal fusion decoder;

[0028] The variable selection network is used to assign attention weights to each input feature;

[0029] The static covariate encoder is used to generate different context vectors based on the static data features after assigning attention weights;

[0030] The gated residual network introduces nonlinear processing into the TFT prediction model;

[0031] The temporal fusion decoder is used to learn the temporal relations present in the training dataset.

[0032] Preferably, the step of assigning attention weights to each input feature in the variable selection network comprises:

[0033] Step 3.1.1: Set is the transformation input of the jth variable at time t, is the flattened vector of all historical inputs at time t;

[0034] Step 3.1.2: By adding Ξ t and an external context vector c s Input into GRN to generate variable selection weights, and then process through the Softmax layer to obtain the vector of variable selection weights;

[0035] Step 3.1.3: At each time step, Input it into its corresponding GRN, apply an additional layer of nonlinear processing to obtain the processed feature vector of variable i;

[0036] Step 3.1.4: The processed feature vector of variable i is shared across all time steps t, weighted according to the variable selection weight of variable i, and an attention weight is assigned to each input feature;

[0037] The vector of variable selection weights is calculated as:

[0038]

[0039] In formula (2), is the vector of variable selection weights, c s is the output of the static covariate encoder;

[0040] The calculation formula for the processed eigenvector of variable i is:

[0041]

[0042] In formula (3), is the processed eigenvector of variable i;

[0043] The expression for weighting the variable selection weight is:

[0044]

[0045] In formula (4), For vector The j-th element of .

[0046] Preferably, the static covariate encoder generates different context vectors according to the static data features after the attention weights are assigned, specifically including:

[0047] According to the output of the static variable selection network, the static covariate encoder uses four independent GRN encoders to generate different context vectors, where the different context vectors include the context selected by the temporal variable (c s ), local processing of temporal features (cc ,c h ) and enriching the temporal features through static information (C e );

[0048] The context of time series variable selection (c s ) is calculated as:

[0049]

[0050] In formula (5), ζ is the output of the static variable selection network.

[0051] Preferably, the gated residual network includes a Dense layer, an ELU, a Gate layer, and an Add&Norm layer, wherein the Dense layer is a fully connected layer, the ELU is an exponential linear unit activation function, the Gate layer is a gated layer, the Add&Norm layer is a residual connection and a standard normalization layer, and the gated residual network introduces nonlinear processing for the TFT prediction model, specifically including:

[0052] When there is a context vector, GRN receives a main input feature vector a and the context vector c output by the static covariate encoder to obtain the output of GRN. When there is no context vector, GRN only regards the context input as 0. During the training of the TFT prediction model, the dropout operation is applied before the Gate layer and the Add&Norm layer, that is, η1 in formula (7), where the feature vector a is the feature processed by the multi-head attention mechanism;

[0053] The output of GRN is expressed as:

[0054] GRN ω (α,c)=LayerNorm(α+GLU ω (η1)) (6);

[0055] η1=W 1,ω η2+b 1,ω (7);

[0056] η2=ELU(W 2,ω α+W 3,ω c+b 2,ω ) (8);

[0057] In formula (6)-formula (8), ELU is the exponential linear unit activation function, is the middle layer, LayerNorm is the standard layer normalization, and ω is the index of weight sharing;

[0058] The expression of ELU is:

[0059] GLU ω(γ)=σ(W 3,ω γ+b 3,ω )⊙(W 4,ω γ+b 4,ω ) (9);

[0060] In formula (9), σ(·) is the sigmoid activation function, is the weight, is the bias, ⊙ is the element-wise Hadamard product, d model is the size of the hidden state.

[0061] Preferably, the step of the temporal fusion decoder learning the temporal relationship existing in the training data set includes:

[0062] Step 3.2.1: Convert the output of the static covariate encoder to The LSTM encoder input into the static covariate encoder will Input to the LSTM decoder to obtain a set of unified time features, which are used as the input of the time series fusion decoder. The input of the time series fusion decoder is φ(t,n)∈{φ(t,-k),...,φ(t,τ max )}, n is the position index;

[0063] Step 3.2.2: Use the context vector c from the static covariate encoder c ,c h , respectively initialize the cell state and hidden state of the first LSTM in the input layer, and use gated skip connections in the input layer to obtain the output time features of the input layer;

[0064] Step 3.2.3: Introduce a static enhancement layer to statically enhance the position index n in the output time feature;

[0065] Step 3.2.4: All statically enhanced temporal features form a matrix Θ(t), Θ(t) = [θ(t, -k), ..., θ(t, τ)] T , in each prediction time step, a multi-head attention mechanism is applied to the temporal features to generate a matrix B(t) = [β(t, -k), ..., β(t, τ max )], and add a gating layer after the multi-head attention layer to facilitate the training process, where the number of heads in the multi-head attention mechanism is m H ;

[0066] The expression of the output time feature of the input layer is:

[0067]

[0068] In formula (10), n∈[-k,τmax ] is the position index;

[0069] The expression of the position index n in the output time feature after static enhancement is:

[0070]

[0071] In formula (11), GRN φ The weights are shared across the layer, c e is the context vector from the static covariate encoder;

[0072] The expression of the multi-head attention mechanism is:

[0073] B(t)=MultiHead(Θ(t),Θ(t),Θ(t)) (12);

[0074] The expression of the gating layer is:

[0075] δ(t,n)=LayerNorm(θ(t,n)+GLU δ (β(t,n))) (13).

[0076] Preferably, outputting the prediction result in step 5 specifically includes:

[0077] The historical features, label sequences, and known auxiliary information are input into the trained TFT prediction model. The variable selection network automatically selects the key features most relevant to the radar working mode prediction task by assigning attention weights to each input feature, and linearly transforms the output of the temporal fusion decoder to generate the prediction result. The known auxiliary information is the temporal relationship existing in the training data set learned by the temporal fusion decoder during the training process.

[0078] The linear transformation expression of the temporal fusion decoder output is:

[0079]

[0080] In formula (13), is the linear coefficient specifying the quantile q, τ∈{1,...,τ max}.

[0081] The beneficial effects of the present invention are:

[0082] 1. The present invention uses the multi-scale radar signal time-frequency diagram feature sequence and radar working mode label sequence extracted based on BiFormer as the prediction basis. BiFormer captures signal features from different scales through its multi-head attention mechanism. The features of each scale correspond to signal changes in different frequency bands, ensuring that the features of the signal are fully characterized within different time ranges.

[0083] 2. The present invention uses a variable selection network to extract key features from the radar signal time-frequency graph feature sequence. The radar signal time-frequency graph feature sequence contains a large amount of possible redundant information. The variable selection network automatically selects the key features most relevant to the radar working mode prediction task by assigning attention weights to each input feature, thereby reducing unnecessary features involved in the calculation.

[0084] 3. The core of the present invention is to use the TFT prediction model to predict the radar working mode. Through the various modules in the TFT prediction model, comprehensive timing information extraction and fusion are performed based on the multi-scale radar signal time-frequency diagram feature sequence and the radar working mode label sequence, and finally the single-step prediction of the radar working mode is realized. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] Figure 1 A flowchart of a radar working mode single-step prediction method based on time series fusion Transformer provided by the present invention;

[0086] Figure 2 A schematic diagram of using the BiFormer model provided by the present invention to extract radar signal time-frequency graph features and labels;

[0087] Figure 3 A schematic diagram of the structure of the TFT prediction model provided by the present invention;

[0088] Figure 4 A schematic diagram of the structure of the gated residual network provided by the present invention;

[0089] Figure 5 A schematic diagram of the structure of the variable selection network provided by the present invention;

[0090] Figure 6 The single-step prediction effect of the method proposed by the present invention under different training degrees and different lengths of working mode sequences provided by the present invention. DETAILED DESCRIPTION

[0091] Combination Figure 1-6 To illustrate this implementation, in this implementation, GRN is a RNN gated recurrent unit, GLU is a gated linear unit, and Dense is a fully connected layer. Figure 1 As shown, the steps of a radar working mode single-step prediction method based on a time series fusion Transformer described in this embodiment include:

[0092] S1: Perform time-frequency transformation on the acquired radar signal to obtain time-frequency graph data, which is used as the original sample data set of the radar working mode;

[0093] S101: Use the WVD formula to calculate the energy distribution of each time segment on the frequency to complete the time-frequency analysis;

[0094] Radar signals collected are IQ signals in the time domain (i.e., in-phase component I and quadrature component Q), which contain the state information of the radar in different working modes. In order to analyze the changing characteristics of these signals in time and frequency, Wigner-Ville distribution (WVD) is used for time-frequency transformation.

[0095] WVD is a high-resolution time-frequency analysis tool suitable for non-stationary signals. Its core idea is to display the distribution of signal energy at different times and frequencies and retain higher time-frequency resolution by combining time and frequency analysis. Its definition formula is as follows:

[0096]

[0097] In formula (1), W x (t,f) is the WVD time-frequency representation of the signal at time t and frequency f, x * (t) is the complex conjugate of the signal, τ is the delay variable in the time domain, which represents the time difference used in the integration process, e -j2πfτ is the frequency component in Fourier transform.

[0098] S102: After the time-frequency analysis is completed, the WVD time-frequency image is obtained in the form of a two-dimensional matrix. In order to adapt to the subsequent model input, these time-frequency images need to be processed into a standard size, so the time-frequency images generated by WVD are adjusted to a size of 224*224*3, where the three channels can represent different frequency bandwidths, time domain features or normalized signal energy;

[0099] S103: Each time-frequency graph is labeled according to its corresponding radar working mode to generate a labeled sample, and the processed time-frequency graph is classified according to the radar working mode to construct a time-frequency graph dataset containing multiple working modes.

[0100] S2: Input the time-frequency diagram in the original sample data set into the BiFormer model to extract and identify multi-scale features. The extracted features and the corresponding radar working mode labels obtained by identification constitute the training data set.

[0101] like Figure 2As shown in the figure, in the initial feature extraction stage, the input time-frequency graph first passes through the Patch embedding module to divide the input image into a group of small image blocks, which are then stacked and convolved along the channel direction to be divided into initial feature blocks of size 56×56×64 and processed by two BiFormer modules. In the intermediate feature extraction stage, the feature map size is reduced to 28×28×128 and 14×14×256 through the Patch merging operation, which are processed by two and eight BiFormer modules respectively. Finally, in the advanced feature extraction stage, the feature map size is reduced to 7×7×512 and then processed by two BiFormer modules. Through this multiple Patch merging operation, the feature map gradually transitions from high resolution to low resolution, and the dimension of the feature gradually increases, providing richer feature representation; among them, the BiFormer module is the core building block based on the dual-level routing attention mechanism, which aims to extract multi-scale features from the input feature map.

[0102] The present invention uses the multi-scale radar signal time-frequency diagram feature sequence and radar working mode label sequence extracted based on BiFormer as the prediction basis. BiFormer captures signal features from different scales through its multi-head attention mechanism. The features of each scale correspond to signal changes in different frequency bands, ensuring that the features of the signal are fully characterized within different time ranges.

[0103] S3: Construct the gated residual network, variable selection network, static covariate encoder and temporal fusion decoder in the TFT prediction model;

[0104] The TFT prediction model constructed in this embodiment is as follows Figure 3 shown.

[0105] S301: Figure 4 As shown, in order to apply nonlinear processing only when necessary, the present embodiment introduces a gated residual network as a basic building block of TFT, and the gated residual network includes a Dense layer, an ELU, a Gate layer, and an Add&Norm layer, wherein the Dense layer is a fully connected layer, the ELU is an exponential linear unit activation function, the Gate layer is a gated layer, and the Add&Norm layer is a residual connection and a standard normalization layer;

[0106] In the presence of a context vector, GRN receives a main input feature vector a and a context vector c output by a static covariate encoder to obtain the output of GRN, where the feature vector a can be the output of an LSTM encoder or an LSTM decoder, or the features output by a variable selection network or processed by a multi-head attention mechanism;

[0107] The output of GRN is expressed as:

[0108] GRN ω (α,c)=LayerNorm(α+GLU ω (η1)) (2);

[0109] η1=W 1,ω η2+b 1,ω (3);

[0110] η2=ELU(W 2,ω α+W 3,ω c+b 2,ω ) (4);

[0111] In formula (2)-formula (4), ELU is the exponential linear unit activation function, is the middle layer, LayerNorm is the standard layer normalization, and ω is the index of weight sharing;

[0112] This implementation uses component gating layers based on gated linear units to provide flexibility, thereby suppressing parts of the architecture that are unnecessary for a specific dataset. As input, the form of GLU is:

[0113]

[0114] In formula (5), σ(·) is the sigmoid activation function, is the weight, is the bias, ⊙ is the element-wise Hadamard product, d model is the size of the hidden state.

[0115] For the case where there is no context vector, the gated residual network only considers the context input as zero, i.e. c = 0 in formula (4). During training, a dropout operation is applied before the gating layer and layer normalization, i.e. η1 in formula (3).

[0116] S302: Although there may be multiple variables available in the multi-scale features of the radar time-frequency graph input to the model, their relevance and specific contribution to the output are usually unknown. In order to select the variables that are most important to the prediction problem and remove unnecessary noise inputs, such as Figure 5 As shown, this implementation introduces a variable selection network, and the specific process is as follows:

[0117] S30201: Set represents the transformed input of the jth variable at time t, where is the flattened vector of all historical inputs at time t. The variable selection weight is calculated by t and an external context vector c sThe input is generated by the gated residual network and then processed by the Softmax layer:

[0118] The vector of variable selection weights is calculated as:

[0119]

[0120] In formula (6), is the vector of variable selection weights, c s is the output of the static covariate encoder;

[0121] S30202: At each time step, by Input to its corresponding gated residual network, and apply an additional layer of nonlinear processing:

[0122] The calculation formula for the processed eigenvector of variable i is:

[0123]

[0124] In formula (7), is the processed eigenvector of variable i;

[0125] S30203: Each variable has its own GRN ξ ( j ), and the weights are shared across all time steps t, weighted according to their variable selection weights, as shown in the following formula:

[0126] The expression for weighting the variable selection weight is:

[0127]

[0128] In formula (8), For vector The j-th element of .

[0129] The present invention uses a variable selection network to extract key features from the feature sequence of radar signal time-frequency graphs. The feature sequence of radar signal time-frequency graphs contains a large amount of possible redundant information. The variable selection network automatically selects the key features most relevant to the radar working mode prediction task by assigning attention weights to each input feature, thereby reducing unnecessary features involved in the calculation.

[0130] S303: In order to construct a complex representation of static metadata, this embodiment uses four independent gated residual network encoders to generate different context vectors, which are then connected to different positions of the temporal fusion decoder, where static variables play an important role in the processing, including the context selected by the temporal variables (c s ), local processing of temporal features (cc ,c h ) and enriching the temporal features through static information (C e ). For example, assuming that ζ is the output of the static variable selection network, the context of temporal variable selection will be according to the formula to encode.

[0131] S304: The temporal fusion decoder constructed in this embodiment uses a series of layers described below to learn the temporal relationship existing in the data set:

[0132] S30401: Enhance locality through sequence-to-sequence layer. Key points in radar working mode feature sequences are usually identified by their relationship with their surrounding values, such as anomalies, change points, or periodic patterns. Therefore, this embodiment uses local context to improve the performance of attention-based architectures by constructing features that utilize pattern information rather than relying solely on point-by-point values. Input to the LSTM encoder, Input to the LSTM decoder, which will generate a set of unified temporal features, which are used as the input of the temporal fusion decoder itself, using φ(t,n)∈{φ(t,-k),...,φ(t,τ max )}, where n is the position index. In order to allow static metadata to affect local processing, c from the static covariate encoder is used c ,c h The context vector initializes the cell state and hidden state of the first LSTM in the input layer, and a gated jump connection is also used on the input layer:

[0133] The expression of the output time feature of the input layer is:

[0134]

[0135] In formula (9), n∈[-k,τ max ] is the position index;

[0136] S30402: Since static covariates usually have a significant impact on temporal dynamics, this implementation introduces a static enhancement layer to enhance temporal features with static metadata. For a given position index n, the form of static enhancement is:

[0137]

[0138] In formula (10), GRN φ The weights are shared across the layer, c e is the context vector from the static covariate encoder;

[0139] S30403: After static enhancement, the self-attention mechanism is applied to the temporal features. All statically enhanced temporal features are first combined into a matrix, namely Θ(t) = [θ(t, -k), ..., θ(t, τ)] T , and apply a multi-head attention mechanism at each prediction time step, where the number of heads in the multi-head attention mechanism is m H , where its form (N = τ max +k+1) is:

[0140] B(t)=MultiHead(Θ(t),Θ(t),Θ(t)) (11);

[0141] S30404: Apply the multi-head attention mechanism to generate B(t) = [β(t, -k), ..., β(t, τ max )], the self-attention layer enables TFT to capture the long-range dependencies in the feature sequence of radar working modes. After the self-attention layer, this embodiment also applies an additional gating layer to facilitate the training process:

[0142] δ(t,n)=LayerNorm(θ(t,n)+GLU δ (β(t,n))) (12).

[0143] S4: inputting the training data set into the TFT prediction model for training to obtain a trained TFT prediction model;

[0144] S5: The received radar signal is also subjected to time-frequency transformation processing and multi-scale feature extraction to obtain corresponding historical features and label sequences;

[0145] S6: Input the historical features and label sequence generated in S5 into the trained TFT prediction model to perform a single-step prediction of the radar working mode. The TFT prediction model outputs a prediction result based on the input feature sequence and known auxiliary information.

[0146] The historical features, label sequences, and known auxiliary information are input into the trained TFT prediction model. The variable selection network automatically selects the key features most relevant to the radar working mode prediction task by assigning attention weights to each input feature, and linearly transforms the output of the temporal fusion decoder to generate the prediction result. The known auxiliary information is the temporal relationship existing in the training data set learned by the temporal fusion decoder during the training process.

[0147] The linear transformation expression of the temporal fusion decoder output is:

[0148]

[0149] In formula (13), is the linear coefficient specifying the quantile q, τ∈{1,...,τ max}.

[0150] In summary, this embodiment uses various modules in the TFT prediction model to extract and fuse comprehensive time series information based on the multi-scale radar signal time-frequency diagram feature sequence and the radar working mode label sequence, and finally realizes the single-step prediction of the radar working mode. The single-step prediction effect of this embodiment under different training degrees and different lengths of working mode sequences is as follows: Figure 6 As shown by Figure 6 It can be seen that the prediction accuracy will increase with the increase of the length of the input task sequence, and the overall prediction result is basically the same as the actual working mode. It can be considered that the method proposed in the present invention is effective and feasible for predicting the working mode of the multi-function radar, which verifies the correctness and effectiveness of the present invention.

[0151] The above is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technician familiar with this profession can make some changes or modify the technical contents disclosed above into equivalent embodiments without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent replacement and improvement made to the above embodiments without departing from the content of the technical solution of the present invention, based on the technical essence of the present invention, within the spirit and principles of the present invention, still fall within the protection scope of the technical solution of the present invention.

Claims

1. A single-step prediction method for radar working mode based on time series fusion Transformer, characterized in that: The steps of the radar working mode single-step prediction method based on time series fusion Transformer include: Step 1: Perform time-frequency transformation on the acquired radar pulse signal to construct the original sample data set of the radar working mode; Step 2: extract and identify multi-scale features of the time-frequency graph in the original sample data set of the radar working mode, and construct a training data set based on the extracted features and the corresponding radar working mode labels obtained by identification; Step 3: Build a TFT prediction model and use the training data set to train the TFT prediction model; Step 4: Perform time-frequency transformation and multi-scale feature extraction on the received radar signal to generate historical features and label sequences; Step 5: Input the generated historical features and label sequences into the trained TFT prediction model to perform single-step prediction of the radar working mode and output the prediction results.

2. According to claim 1, a radar working mode single-step prediction method based on time series fusion Transformer is characterized in that: The steps of constructing the radar working mode original sample data set in step 1 include: Step 1.1: Use the WVD formula to calculate the energy distribution of the acquired radar pulse signal in frequency at each time segment, complete the time-frequency analysis, and obtain a WVD time-frequency image in the form of a two-dimensional matrix, where the radar pulse signal is an IQ signal in the time domain, I is the in-phase component, Q is the orthogonal component, and the IQ signal contains the status information of the radar in different working modes; Step 1.2: Adjust the WVD time-frequency image in the form of a two-dimensional matrix to a size of 224*224*3 to complete the WVD time-frequency image standardization process, where the three channels have different frequency bandwidths and time domain features; Step 1.3: The WVD time-frequency image is annotated according to the corresponding radar working mode to generate a labeled sample. The standardized WVD time-frequency image is classified according to the radar working model to construct the radar working mode original sample data set; The WVD formula is: In formula (1), W x (t,f) is the WVD time-frequency representation of the signal at time t and frequency f, x * (t) is the complex conjugate of the signal, τ is the delay variable in the time domain, which represents the time difference used in the integration process, e -j2πfτ is the frequency component in Fourier transform.

3. According to claim 1, a radar working mode single-step prediction method based on time series fusion Transformer is characterized in that: In step 2, multi-scale feature extraction and recognition of the time-frequency graph in the original sample data set of the radar working mode specifically include: The original sample dataset of radar working mode is input into the multi-scale feature extraction architecture based on the BiFormer model to extract multi-scale features. The extracted multi-scale features are input into the classification module composed of layer normalization layer, average pooling layer, tensor flattening layer and classification head to obtain the corresponding radar working mode label.

4. According to claim 3, a radar working mode single-step prediction method based on time series fusion Transformer is characterized in that: The multi-scale feature extraction architecture based on the BiFormer model includes a Patch embedding module, a BiFormer module, and a Patch merging module; The Patch embedding module is used to divide the input time-frequency graph into different sizes; The BiFormer module is a core building block based on a two-level routing attention mechanism, used to extract multi-scale features from the input feature map; The Patch merging module is used to reduce the size of the feature map processed by the BiFormer module. Extracting multi-scale features specifically includes: In the initial feature extraction stage, the Patch embedding module divides the input time-frequency map into a set of small image patches, stacks and convolves the time-frequency map along the channel direction, divides it into initial feature patches of size 56×56×64, and processes them through two BiFormer modules; In the intermediate feature extraction stage, the feature map sizes processed by the two BiFormer modules are reduced to 28×28×128 and 14×14×256 by the Patch merging module, the feature map with a size of 28×28×128 is processed by two BiFormer modules, and the feature map with a size of 14×14×256 is processed by eight BiFormer modules; In the high-level feature extraction stage, the feature map sizes processed by two BiFormer modules and eight BiFormer modules are uniformly reduced to 7×7×512 through the Patch merging module and processed by two BiFormer modules to complete multi-size feature extraction.

5. According to claim 1, a radar working mode single-step prediction method based on time series fusion Transformer is characterized in that: The TFT prediction model constructed in step 3 includes a variable selection network, a static covariate encoder, a gated residual network, and a temporal fusion decoder; The variable selection network is used to assign attention weights to each input feature; The static covariate encoder is used to generate different context vectors according to the static data features after the attention weights are assigned; The gated residual network introduces nonlinear processing into the TFT prediction model; The temporal fusion decoder is used to learn the temporal relationship existing in the training data set.

6. According to claim 5, the radar working mode single-step prediction method based on time series fusion Transformer is characterized in that: The steps for the variable selection network to assign attention weights to each input feature include: Step 3.1.1: Set is the transformation input of the jth variable at time t, is the flattened vector of all historical inputs at time t; Step 3.1.2: By adding Ξ t and an external context vector c s Input into GRN to generate variable selection weights, and then process through the Softmax layer to obtain the vector of variable selection weights; Step 3.1.3: At each time step, Input it into its corresponding GRN, apply an additional layer of nonlinear processing to obtain the processed feature vector of variable i; Step 3.1.4: The processed feature vector of variable i is shared across all time steps t, weighted according to the variable selection weight of variable i, and an attention weight is assigned to each input feature; The vector of variable selection weights is calculated as: In formula (2), is the vector of variable selection weights, c s is the output of the static covariate encoder; The calculation formula for the processed eigenvector of variable i is: In formula (3), is the processed eigenvector of variable i; The expression for weighting the variable selection weight is: In formula (4), For vector The j-th element of .

7. According to claim 5, the radar working mode single-step prediction method based on time series fusion Transformer is characterized in that: The static covariate encoder generates different context vectors based on the static data features after assigning attention weights, including: According to the output of the static variable selection network, the static covariate encoder uses four independent GRN encoders to generate different context vectors, where the different context vectors include the context selected by the temporal variable (c s ), local processing of temporal features (c c ,c h ) and enriching the temporal features through static information (C e ); The context of time series variable selection (c s ) is calculated as: In formula (5), ζ is the output of the static variable selection network.

8. According to claim 5, the radar working mode single-step prediction method based on time series fusion Transformer is characterized in that: The gated residual network includes Dense layer, ELU, Gate layer and Add&Norm layer, where the Dense layer is a fully connected layer, ELU is an exponential linear unit activation function, the Gate layer is a gated layer, the Add&Norm layer is a residual connection and a standard normalization layer. The gated residual network introduces nonlinear processing for the TFT prediction model, including: When there is a context vector, GRN receives a main input feature vector a and the context vector c output by the static covariate encoder to obtain the output of GRN. When there is no context vector, GRN only regards the context input as 0. During the training of the TFT prediction model, the dropout operation is applied before the Gate layer and the Add&Norm layer, that is, η1 in formula (7), where the feature vector a is the feature processed by the multi-head attention mechanism; The output of GRN is expressed as: GRN ω (α,c)=LayerNorm(α+GLU ω (h1)) (6); η1=W 1,ω η2+b 1,ω (7); η2=ELU(W 2,ω α+W 3,ω c+b 2,ω )(8); In formula (6)-formula (8), ELU is the exponential linear unit activation function, is the middle layer, LayerNorm is the standard layer normalization, and ω is the index of weight sharing; The expression of ELU is: GLU ω (c)=σ(W 3,ω c+b 3,ω )⊙(W 4,ω c+b 4,ω ) (9); In formula (9), σ(·) is the sigmoid activation function, is the weight, is the bias, ⊙ is the element-wise Hadamard product, d model is the size of the hidden state.

9. According to claim 5, the radar working mode single-step prediction method based on time series fusion Transformer is characterized in that: The steps for the temporal fusion decoder to learn the temporal relationship existing in the training dataset include: Step 3.2.1: Convert the output of the static covariate encoder to The LSTM encoder input into the static covariate encoder will Input to the LSTM decoder to obtain a set of unified time features, which are used as the input of the time series fusion decoder. The input of the time series fusion decoder is φ(t,n)∈{φ(t,-k),...,φ(t,τ max )}, n is the position index; Step 3.2.2: Use the context vector c from the static covariate encoder c ,c h , respectively initialize the cell state and hidden state of the first LSTM in the input layer, and use gated skip connections in the input layer to obtain the output time features of the input layer; Step 3.2.3: Introduce a static enhancement layer to statically enhance the position index n in the output time feature; Step 3.2.4: All statically enhanced temporal features form a matrix Θ(t), Θ(t) = [θ(t, -k), ..., θ(t, τ)] T , in each prediction time step, a multi-head attention mechanism is applied to the temporal features to generate a matrix B(t) = [β(t, -k), ..., β(t, τ max )], and add a gating layer after the multi-head attention layer to facilitate the training process, where the number of heads in the multi-head attention mechanism is m H ; The expression of the output time feature of the input layer is: In formula (10), n∈[-k,τ max ] is the position index; The expression of the position index n in the output time feature after static enhancement is: In formula (11), GRN φ The weights are shared across the layer, c e is the context vector from the static covariate encoder; The expression of the multi-head attention mechanism is: B(t)=MultiHead(Θ(t),Θ(t),Θ(t)) (12); The expression of the gating layer is: δ(t,n)=LayerNorm(θ(t,n)+GLU δ (β(t,n)))(13).

10. The radar working mode single-step prediction method based on time series fusion Transformer according to claim 1 is characterized in that: The prediction results output in step 5 specifically include: The historical features, label sequences, and known auxiliary information are input into the trained TFT prediction model. The variable selection network automatically selects the key features most relevant to the radar working mode prediction task by assigning attention weights to each input feature, and linearly transforms the output of the temporal fusion decoder to generate the prediction result. The known auxiliary information is the temporal relationship existing in the training data set learned by the temporal fusion decoder during the training process. The linear transformation expression of the temporal fusion decoder output is: In formula (13), is the linear coefficient specifying the quantile q, τ∈{1,...,τ max }.

Images