A radar working mode single-step prediction method based on time sequence fusion Transformer

Abstract

The application provides a radar working mode single-step prediction method based on a time sequence fusion Transformer, belongs to the technical field of radar signal processing, and solves the problems of low prediction accuracy and poor prediction function robustness in traditional radar working mode prediction. The specific steps comprise the following: performing transform domain analysis on the obtained radar pulse signals to construct a radar working mode original sample data set; performing multi-scale feature extraction and identification on the time-frequency graph in the radar working mode original sample data set, constructing a training data set based on the extracted features and the corresponding radar working mode labels obtained through identification; 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 a historical feature and a label sequence; inputting the generated historical feature and label sequence into the trained TFT prediction model to perform single-step prediction of the radar working mode, and outputting a prediction result.

Description

TECHNICAL FIELD

[0001] The present application relates to a radar working mode single-step prediction method based on temporal fusion Transformer, belonging to the technical field of radar signal processing. BACKGROUND

[0002] With the rapid development of phased array radar technology, its flexible beam pointing, waveform modulation and multi-task processing capability greatly improve the adaptability and complexity of the radar system, which brings great technical challenges to radar reconnaissance and countermeasures. Radar working mode prediction is one of the important technologies in the field of electronic warfare, which aims to predict the future working state of the radar in advance through the analysis of the intercepted radar pulse signal. The radar working mode prediction algorithm needs to identify the historical working mode, and based on the time sequence characteristics of the pulse signal, the internal law of the radar behavior is mined, so as to realize more accurate and reliable working mode prediction.

[0003] Temporal Fusion Transformer (TFT) is an architecture designed for time series prediction tasks, which has the ability to capture time series dependencies, especially 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 phased processing, dynamically select the most prominent and core features, and use global and local time dependencies to predict the target sequence. SUMMARY

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

[0005] Step 1: Transform domain analysis is performed on the acquired radar pulse signal to construct a radar working mode original sample data set;

[0006] Step 2: Multi-scale feature extraction and identification are performed on the time-frequency graph in the radar working mode original sample data set, and a training data set is constructed based on the extracted features and the corresponding radar working mode labels obtained by identification;

[0007] Step 3: Construct a TFT prediction model and train the TFT prediction model using the training data set;

[0008] Step 4: Transform domain analysis and multi-scale feature extraction are performed on the received radar signal to generate a historical feature and label sequence;

[0009] Step 5: The generated historical features and label sequence are input into the trained TFT prediction model to perform single-step prediction of the radar working mode, and the prediction result is output.

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

[0011] Step 1.1: Calculate the energy distribution of the acquired radar pulse signal in each time slice in frequency using the WVD formula, complete the time-frequency analysis, and obtain a two-dimensional matrix form WVD time-frequency image, wherein the radar pulse signal is an IQ signal in the time domain, I is the in-phase component, Q is the quadrature component, and the IQ signal contains the state information of the radar in different working modes;

[0012] Step 1.2: Adjust the two-dimensional matrix form WVD time-frequency image to the size of 224*224*3, complete the WVD time-frequency image standardization processing, wherein the three channels are different frequency bandwidth and time domain features;

[0013] Step 1.3: The WVD time-frequency image is labeled according to the corresponding radar working mode to generate a sample with a label, the standardized WVD time-frequency image is classified according to the radar working model, and a radar working mode original sample data set is constructed;

[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, represents the time difference used in the integration process, e -j2πfτ is the frequency component in the Fourier transform.

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

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

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

[0020] The patch embedding module is used for size division of the input time-frequency graph;

[0021] The BiFormer module is a core construction unit based on a two-level routing attention mechanism, and is used for extracting multi-scale features from the input feature map;

[0022] The patch merging module is used for size reduction of the feature map processed by the BiFormer module.

[0023] The step of extracting multi-scale features specifically includes:

[0024] In the initial feature extraction stage, the patch embedding module divides the input time-frequency graph into a group of small image blocks, stacks and convolves the time-frequency graph along the channel direction, divides the time-frequency graph into initial feature blocks with a size of 56x56x64, and processes the initial feature blocks through two BiFormer modules;

[0025] In the intermediate feature extraction stage, the patch merging module reduces the sizes of the feature maps processed by the two BiFormer modules to 28x28x128 and 14x14x256, processes the feature map with a size of 28x28x128 through two BiFormer modules, and processes the feature map with a size of 14x14x256 through eight BiFormer modules;

[0026] In the high-level feature extraction stage, the patch merging module uniformly reduces the sizes of the feature maps processed by the two BiFormer modules and the eight BiFormer modules to 7x7x512, and processes the feature maps through two BiFormer modules to complete multi-scale 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 time series 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 according to the static data features after the attention weights are assigned;

[0030] The gated residual network introduces a nonlinear process for the TFT prediction model;

[0031] The time series fusion decoder is used to learn the time series relationship existing in the training data set.

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

[0033] Step 3.1.1: Let the transformed input for the jth variable at time t, the flattened vector of all historical inputs for time t;

[0034] Step 3.1.2: the transformed input for the jth variable at time t, t and an external context vector c s is input into the GRN to generate variable selection weights, and the variable selection weights are processed through a Softmax layer to obtain a vector of variable selection weights;

[0035] Step 3.1.3: at each time step, the transformed input for the jth variable at time t, is input into its corresponding GRN to apply an additional layer of non-linear processing to obtain a processed feature vector for the variable i;

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

[0037] The formula for calculating the vector of variable selection weights is:

[0038]

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

[0040] The formula for calculating the processed feature vector for the variable i is:

[0041]

[0042] In formula (3), is the processed feature vector for the variable i;

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

[0044]

[0045] In formula (4), is the jth element of the vector .

[0046] Preferably, the static covariate encoder generates different context vectors from the static data features after assigning attention weights, 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 of the temporal variable selection (c s ), the local processing of the temporal features (cc ,c h ) and enrich the timing features (C e ) by static information;

[0048] The formula for calculating the context (c s ) of the timing variable selection is as follows:

[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 gating layer, and the Add&Norm layer is a residual connection and a standard normalization layer. The gated residual network introduces a nonlinear processing for the TFT prediction model, which specifically includes:

[0052] In the presence of a context vector, the GRN receives a main input feature vector a and a context vector c output by the static covariant encoder to obtain the output of the GRN. In the absence of a context vector, the GRN only considers the input of the context as 0. In the training process of the TFT prediction model, a dropout operation is applied before the Gate layer and the Add&Norm layer, that is, η1 in formula (7), wherein the feature vector a is the feature processed by the multi-head attention mechanism;

[0053] The expression of the output of the GRN is as follows:

[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 an exponential linear unit activation function, is an intermediate layer, LayerNorm is a standard layer normalization, and ω is a weight-shared index;

[0058] The expression of the ELU is as follows:

[0059] GLU ω(γ) = σ(W 3,ω γ + b 3,ω ) (8). 4,ω γ + b 4,ω ) (9).

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

[0061] Preferably, the step of learning and training the time sequence relationship existing in the training data set by the time sequence fusion decoder comprises:

[0062] Step 3.2.1: inputting the static covariate encoder output to the LSTM encoder in the static covariate encoder, inputting to the LSTM decoder to obtain a set of unified time features, and taking the unified time features as the input of the time sequence fusion decoder, wherein the input of the time sequence fusion decoder is φ(t, n) ∈ {φ(t, -k),..., φ(t, τ max )} and n is a position index;

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

[0064] Step 3.2.3: introducing a static enhancement layer to perform static enhancement on the position index n in the output time features;

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

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

[0067]

[0068] In formula (10), n ∈ [-k, τmax ] is a 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 weight is shared in the whole 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(Theta(t), Theta(t), Theta(t)) (12).

[0074] The expression of the gating layer is:

[0075] delta(t, n) = LayerNorm(theta(t, n) + GLU δ (beta(t, n)) (13).

[0076] Preferably, the output prediction result in step 5 specifically comprises:

[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 related to the radar working mode prediction task by assigning attention weights to each input feature, and a linear transformation is performed on the time sequence fusion decoder output to generate a prediction result, wherein the known auxiliary information is the time sequence relationship learned by the time sequence fusion decoder in the training data set during the training process.

[0078] The expression of the linear transformation performed on the time sequence fusion decoder output is:

[0079]

[0080] In formula (13), tau is the linear coefficient of the specified quantile q, tau element of {1,..., tau max}.

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

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

[0083] 2、The application uses a variable selection network to extract key features in 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 filters out the key features most relevant to the radar working mode prediction task by assigning attention weights to each input feature, thereby reducing unnecessary features from participating in the calculation.

[0084] 3、The core of the application is to predict the radar working mode by using the TFT prediction model. Through each module in the TFT prediction model, comprehensive time sequence information extraction and fusion are carried out based on the multi-scale radar signal time-frequency graph 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 DRAWINGS

[0085] Figure 1 A flowchart of a radar working mode single-step prediction method based on time sequence fusion Transformer provided by the application is provided.

[0086] Figure 2 An illustration of using a BiFormer model to extract radar signal time-frequency graph features and labels provided by the application is provided.

[0087] Figure 3 A structural diagram of a TFT prediction model provided by the application is provided.

[0088] Figure 4 A structural diagram of a gated residual network provided by the application is provided.

[0089] Figure 5 A structural diagram of a variable selection network provided by the application is provided.

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

[0091] In combination Figures 1-6 In this embodiment, GRN is an RNN gated recurrent unit, GLU is a gated linear unit, and Dense is a fully connected layer, as shown in Figure 1 The steps of a radar working mode single-step prediction method based on time sequence 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 radar working mode original sample data set;

[0093] S101: Calculate the energy distribution of each time slice in frequency using the WVD formula, complete time-frequency analysis;

[0094] The radar signal collected is the IQ signal in the time domain (i.e. in-phase component I and quadrature component Q), which contains the state information of the radar in different working modes. In order to analyze the change 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 show the energy distribution of signals at different times and frequencies through joint time and frequency analysis, and to retain higher time-frequency resolution. 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, representing the time difference used in the integration process, e -j2πfτ is the frequency component in the Fourier transform.

[0098] S102: After time-frequency analysis, the WVD time-frequency image obtained is 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. Therefore, the time-frequency image generated by WVD is adjusted to a size of 224*224*3, where the three channels can represent different frequency bandwidths, time domain features or normalized signal energies.

[0099] S103: Each time-frequency image will be labeled according to its corresponding radar working mode to generate a sample with labels. The processed time-frequency images are classified according to the radar working mode to construct a time-frequency image dataset containing multiple working modes.

[0100] S2: Input the time-frequency image in the original sample dataset into the BiFormer model for multi-scale feature extraction and recognition. The extracted features and the corresponding radar working mode labels obtained by recognition constitute the training dataset.

[0101] For example Figure 2As shown, 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, and then stack and convolve them along the channel direction to divide them into initial feature blocks with a size of 56x56x64, and then pass them through two BiFormer modules. In the intermediate feature extraction stage, the feature map size is reduced to 28x28x128 and 14x14x256 through the Patch merging operation, and then processed by two and eight BiFormer modules, respectively. Finally, in the advanced feature extraction stage, the feature map size is reduced to 7x7x512, 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 to provide richer feature representation. The BiFormer module is a core building unit based on a two-level routing attention mechanism, which aims to extract multi-scale features from the input feature map.

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

[0103] S3: Constructing a gated residual network, a variable selection network, a static covariate encoder, and a time sequence fusion decoder in the TFT prediction model;

[0104] The TFT prediction model constructed by the present embodiment is as shown in Figure 3 .

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

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

[0107] The expression of the output of the GRN is:

[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 equations (2)-(4), ELU is the exponential linear unit activation function, is the intermediate layer, LayerNorm is the standard layer normalization, and ω is the index of weight sharing.

[0112] The present embodiment employs a component gating layer based on the gated linear unit to provide flexibility to suppress unnecessary architecture parts for a specific dataset. is the input, and the form of GLU is:

[0113]

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

[0115] For the case without context vectors, the gated residual network only considers the context input as zero, i.e., c = 0 in equation (4). During the training process, a dropout operation is applied before the gating layer and the layer normalization, i.e., η1 in equation (3).

[0116] S302: Although there can be multiple variables available in the radar time-frequency map multi-scale features input by the model, their relevance and specific contribution to the output are usually unknown. To select the most important variables for the prediction problem and remove unnecessary noise inputs, the present embodiment introduces a variable selection network, as shown in the following: Figure 5

[0117] S30201: Let denote 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 obtained by combining Ξ t and an external context vector c s ​The input is generated into a gated residual network, and then processed through a Softmax layer:

[0118] The calculation formula of the vector of variable selection weights is:

[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, each is input into its corresponding gated residual network, and an additional layer of nonlinear processing is applied:

[0122] The calculation formula of the processed feature vector of variable i is:

[0123]

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

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

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

[0127]

[0128] In formula (8), is the jth element of the vector .

[0129] The present application uses a variable selection network to extract key features in 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 filters out the key features most relevant to the radar working mode prediction task by assigning attention weights to each input feature, thereby reducing unnecessary features from participating in the calculation.

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

[0131] S304: The temporal fusion decoder constructed by the 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 layers, key points in radar operating mode feature sequences are usually identified by their relationship with surrounding values, such as abnormal points, change points or periodic patterns, therefore, the embodiment utilizes local context, by constructing features that utilize pattern information rather than just relying on point-by-point values, can improve the performance of attention-based architecture, input to the LSTM encoder, and to the LSTM decoder, which generates a set of unified temporal features, which are used as inputs to the temporal fusion decoder itself, denoted by φ(t,n)∈{φ(t,-k),...,φ(t,τ max}, where n is the position index. In order to let static metadata affect local processing, use c c h context vectors to initialize the cell state and hidden state of the first LSTM of the input layer, respectively, and also use gated skip connections on the input layer:

[0133] The expression of the output temporal features 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 important influence on time dynamics, the embodiment introduces a static enhancement layer to enhance time features with static metadata, and the form of static enhancement for a given position index n is:

[0137]

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

[0139] ​​S30403: After static enhancement, a self-attention mechanism is applied to the temporal features. All the statically enhanced temporal features are first combined into a matrix, i.e., Θ(t)=[θ(t,-k),...,θ(t,τ)] T A multi-head attention mechanism is applied at each prediction time step, where the number of multi-head attention mechanisms 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: After applying a multi-head attention mechanism, B(t) = [β(t,-k),...,β(t,τ)] is generated. max The self-attention layer enables the TFT to capture long-range dependencies in the radar operating mode feature sequence. Following the self-attention layer, this implementation 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: Input the training dataset into the TFT prediction model for training to obtain the trained TFT prediction model;

[0144] S5: The received radar signals are also processed by time-frequency transformation and multi-scale feature extraction to obtain the corresponding historical features and label sequences.

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

[0146] 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. The output of the temporal fusion decoder is linearly transformed to generate the prediction result. The known auxiliary information is the temporal relationship in the training dataset learned by the temporal fusion decoder during the training process.

[0147] The expression for the linear transformation of the output of the time-series fusion decoder is:

[0148]

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

[0150] In summary, the present embodiment realizes single-step prediction of radar working mode through each module in the TFT prediction model, based on comprehensive time sequence information extraction and fusion of multi-scale radar signal time-frequency graph feature sequence and radar working mode label sequence. The single-step prediction effect of the present embodiment under different training degrees and different length working mode sequences is shown in Table 1, and the prediction accuracy is shown in Table 2. Figure 6 As can be seen from Tables 1 and 2, the prediction accuracy increases 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, so it can be considered that the method proposed in the present embodiment is effective and feasible for multi-functional radar working mode prediction, and the correctness and effectiveness of the present embodiment are verified. Figure 6

[0151] The above is only a preferred embodiment of the present application, and does not limit the present application in any form. Although the present application has been disclosed as above with a preferred embodiment, it is not intended to limit the present application. Any skilled person in the art can make some changes or modifications to the above disclosed technical content without departing from the scope of the technical solution of the present application, and the equivalent embodiments with equivalent changes are equivalent. Any simple modification, equivalent replacement and improvement of the above embodiments, as long as it does not deviate from the technical solution of the present application, belongs to the protection scope of the present application.​

Claims

1. A radar working mode single-step prediction method based on timing fusion Transformer, characterized in that, The steps of the radar working mode single-step prediction method based on the time sequence fusion Transformer include: Step 1: performing time-frequency transformation on the obtained radar pulse signal to construct a radar working mode original sample data set; Step 2: performing multi-scale feature extraction and identification on the time-frequency graph in the radar working mode original sample data set, and constructing a training data set based on the extracted features and the corresponding radar working mode labels obtained by identification; The multi-scale feature extraction and identification of the time-frequency graph in the radar working mode original sample data set in step 2 specifically includes: inputting the radar working mode original sample data set into a multi-scale feature extraction architecture based on a BiFormer model to extract multi-scale features, and inputting the extracted multi-scale features into a classification module composed of a layer normalization layer, an average pooling layer, a tensor flattening layer and a classification head to obtain the corresponding radar working mode label; 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 by size; The BiFormer module is a core construction unit based on a double-level routing attention mechanism, which is 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; The multi-scale feature extraction specifically includes: In the initial feature extraction stage, the Patch embedding module divides the input time-frequency graph into a group of small image blocks, stacks and convolves the time-frequency graph along the channel direction, divides it into initial feature blocks with a size of 56x56x64, and processes it through two BiFormer modules; In the intermediate feature extraction stage, the Patch merging module reduces the size of the feature maps processed by the two BiFormer modules to 28x28x128 and 14x14x256, processes the feature map with a size of 28x28x128 through two BiFormer modules, and processes the feature map with a size of 14x14x256 through eight BiFormer modules; In the high-level feature extraction stage, the Patch merging module uniformly reduces the sizes of the feature maps processed by the two BiFormer modules and the eight BiFormer modules to 7x7x512, and processes them through two BiFormer modules to complete multi-scale feature extraction; Step 3: constructing a TFT prediction model and training the TFT prediction model using the training data set; The TFT prediction model constructed in step 3 includes a variable selection network, a static covariate encoder, a gated residual network and a time sequence 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 a nonlinear process for the TFT prediction model; The time sequence fusion decoder is used to learn the time sequence relationship existing in the training data set; Step 4: Time-frequency transformation and multi-scale feature extraction are performed on the received radar signal to generate historical features and label sequences; Step 5: The generated historical features and label sequences are input into the trained TFT prediction model to perform single-step prediction of the radar working mode, and the prediction result is output.

2. The radar working mode single-step prediction method based on timing fusion Transformer according to claim 1, characterized in that, The steps of constructing the radar working mode original sample data set in step 1 include: Step 1.1: Calculate the energy distribution of the acquired radar pulse signal in each time slice in frequency using the WVD formula, complete the time-frequency analysis, and obtain the WVD time-frequency image in the form of a two-dimensional matrix, wherein the radar pulse signal is an IQ signal in the time domain, I is the in-phase component, Q is the quadrature component, and the IQ signal contains the state 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, and complete the WVD time-frequency image standardization processing, wherein the three channels are different frequency bandwidth and time domain features; Step 1.3: The WVD time-frequency image is labeled according to the corresponding radar working mode to generate a sample with a label, the standardized WVD time-frequency image is classified according to the radar working model, and a radar working mode original sample data set is constructed; The WVD formula is: （1）； In equation (1), is the WVD time-frequency representation of the signal at time and frequency is the complex conjugate of the signal, is the delay variable in the time domain, representing the time difference used in the integration process, is the frequency component in the Fourier transform.​ 3. The radar working mode single-step prediction method based on timing fusion Transformer according to claim 1, characterized in that, The variable selection network assigns attention weights to each input feature, which includes: Step 3.1.1: Let For the first j Variables in time t Transformed input, For time t Flattened vectors of all historical inputs; Step 3.1.2: Generating variable selection weights by inputting and an external context vector into the GRN, and processing through a Softmax layer to obtain a vector of variable selection weights; Step 3.1.3: At each time step, the input to its corresponding GRN is applied with an additional layer of non-linear processing to obtain the processed feature vector of variable i ;​ Step 3.1.4: Variables i The processed feature vectors at all time steps t Shared on top, based on variables i The variables are selected and weighted, and attention weights are assigned to each input feature; The calculation formula of the vector of variable selection weights is: （2）； In equation (2), is a vector of variable selection weights, is the output of the static covariate encoder; Variable i The formula for calculating the post-processing feature vector of the variable is: （3）； In equation (3), is a processed feature vector for the variable i is a processed feature vector for the variable The expression for weighting the variable selection weights is: （4）； In equation (4), is the jth element of the vector is the jth element of the vector 4. The radar working mode single-step prediction method based on timing fusion Transformer according to claim 1, characterized in that, The static covariate encoder generates different context vectors according to the static data features after assigning attention weights, which includes: According to the static variable selection network, the static covariate encoder uses four independent GRN encoders to generate different context vectors, wherein the different context vectors include a context of a time series variable selection , local processing of time series features , and enriching time series features by static information ; Context for timing variable selection The formula for calculating is: （5）； In equation (5), The output of the network is chosen as a static variable.

5. The radar working mode single-step prediction method based on timing fusion Transformer according to claim 1, characterized in that, The gated residual network comprises a Dense layer, , a Gate layer and an Add&Norm layer, wherein the Dense layer is a fully connected layer, is an exponential linear unit activation function, the Gate layer is a gating layer, the Add&Norm layer is a residual connection and a standard normalization layer, and the gated residual network introduces a non-linear processing for the TFT prediction model, specifically comprising: In the presence of context vectors, the GRN receives a main input feature vector a and the context vector output by the static covariate encoder c The output of the GRN is obtained, in the absence of context vectors, the GRN only considers the input of the context as 0, and in the TFT prediction model training process, dropout operation is applied before the Gate layer and the Add&Norm layer, that is, formula (7) where the feature vector a is the feature processed by the multi-head attention mechanism; The output expression of the GRN is: （6）； （7）； （8）； in Equations (6) - (8), is an exponential linear unit activation function, is an intermediate layer, is a standard layer normalization, is an index for weight sharing; The expression is: （9）； In equation (9), is a sigmoid activation function, is a weight, is a bias, is an element-wise Hadamard product, is the size of the hidden state.

6. The radar working mode single-step prediction method based on timing fusion Transformer according to claim 1, characterized in that, The steps of the time sequence fusion decoder learning the time sequence relationship existing in the training data set include: Step 3.2.1: encode the output of the static covariate encoder into an LSTM encoder, which is input into the LSTM decoder to obtain a set of unified temporal features, which are input into the temporal fusion decoder, where the input of the temporal fusion decoder is , , n is a position index; Step 3.2.2: Using the context vector from the static covariate encoder The cell state and hidden state of the first LSTM of the input layer are initialized respectively, and the gated skip connection is adopted in the input layer to obtain the output time features of the input layer. Step 3.2.3: Introduce static enhancement layer to the position index in the output temporal profile n Perform static enhancement; Step 3.2.4: All the time features that have been statically enhanced form a matrix , At each prediction time step, a multi-head attention mechanism is applied to the time features to generate a matrix , and a gating layer is added after the multi-head attention layer to facilitate the training process, where the number of heads in the multi-head attention mechanism is ; The output time feature expression of the input layer is: （10）； In equation (10), is a position index; Position index in output temporal features after static augmentation n The expression is: （11）； In equation (11), the weights of the are shared across the layers, is the context vector from the static covariate encoder; The expression of the multi-head attention mechanism is: （12）； The expression of the gating layer is: （13）。 7. The radar working mode single-step prediction method based on timing fusion Transformer according to claim 1, characterized in that, The output prediction result in step 5 includes: The historical features, label sequences, and known auxiliary information are input into the trained TFT prediction model, the variable selection network assigns attention weights to each input feature, automatically selects the key features most relevant to the radar working mode prediction task, and performs linear transformation on the time sequence fusion decoder output to generate the prediction result, wherein the known auxiliary information is the time sequence relationship learned by the time sequence fusion decoder in the training data set during training; The expression of the linear transformation of the time sequence fusion decoder output is: （13）； In equation (13), is a linear coefficient that specifies the quantile . .

Images