Characterization enhancement-based time sequence basic model construction method and electromagnetic recognition system

Abstract

The invention discloses a time sequence basic model construction method based on representation enhancement and an electromagnetic recognition system. The method comprises the following steps: constructing a time sequence basic model; dividing an input time sequence into a plurality of blocks, and performing linear mapping and one-dimensional convolution operation on each block; key point selection and nonlinear enhancement are carried out on initial block embedding, and initial enhanced block embedding is obtained; the method comprises the following steps of: inputting data into an encoder, extracting key points by applying Fourier transform and inverse Fourier transform to block embedding, carrying out cross attention operation on the block embedding and key point embedding, carrying out self-attention operation on the block embedding, and carrying out operation of a general transformer module; and inputting the time sequence characteristics extracted by the multi-layer transformer module into a linear layer, and identifying the type of the electromagnetic signal. According to the method, fine-grained time information is effectively reserved, the effectiveness of time sequence feature extraction by the time sequence basic model is remarkably enhanced, and meanwhile, the accuracy of electromagnetic signal type recognition is greatly improved.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a method for constructing a time series basic model based on representation enhancement and an electromagnetic recognition system. Background Art

[0002] In the digital age, time series data is widely present in fields such as communications, radar, and medical treatment, and is of great significance. Deep learning, with its advantages in nonlinear modeling and automatic feature extraction, can mine patterns from massive amounts of complex data and provide support for time series related tasks (such as communication signal analysis, radar target recognition, and medical physiological signal monitoring). Accurate time series analysis is crucial for decision-making in various industries. It can help optimize signal transmission strategies in the communications field, accurately detect and track targets in radar systems, and achieve efficient disease diagnosis and monitoring in the medical industry. However, existing time series processing methods have shortcomings. Due to insufficient block representation learning, the existing sequence block-based methods have poor accuracy in time series feature extraction and signal type recognition in these fields using the constructed time series basic models. Summary of the Invention

[0003] The purpose of this invention is to provide a method for constructing a time-series-based model based on representation enhancement and an electromagnetic recognition system. To address the problem that existing methods ignore fine-grained information in block annotation learning, a block annotation enhancement mechanism is proposed. This method screens key points within a block and incorporates this key point information into the block features, improving the local perception of the block and ultimately enhancing the model's recognition performance.

[0004] The technical solution for achieving the purpose of the present invention is as follows: In a first aspect, the present invention provides a method for constructing a time series basic model based on representation enhancement, comprising the following steps:

[0005] Step 1: Using electromagnetic signal data containing multi-dimensional time series features as original samples, a time series basic model is constructed;

[0006] Step 2: Divide the input time series into multiple blocks, perform linear mapping and one-dimensional convolution on each block, and obtain the initial block embedding and point embedding;

[0007] Step 3: Before inputting to the encoder, key points of the initial block embedding are selected and nonlinearly enhanced to obtain the initial enhanced block embedding;

[0008] Step 4: After inputting into the encoder, the block embedding is subjected to Fourier transform and inverse Fourier transform to select key points, the block embedding and key point embedding are subjected to cross attention operation, and then the block embedding is subjected to self-attention operation, and then the transformer module operation is performed;

[0009] In step 5, the time series features extracted by the multi-layer transformer module are input into the linear layer for electromagnetic signal type recognition.

[0010] In a second aspect, the present invention provides a time-series based model electromagnetic signal recognition system based on characterization enhancement, for implementing the method described in the first aspect, the system comprising:

[0011] The first module is used to construct a time series basic model using electromagnetic signal data containing multi-dimensional time series features as original samples;

[0012] The second module is used to divide the input time series into multiple blocks, perform linear mapping and one-dimensional convolution on each block, and obtain the initial block embedding and point embedding;

[0013] The third module is used to select key points of the initial block embedding and perform nonlinear enhancement on it before inputting it into the encoder to obtain the initial enhanced block embedding;

[0014] The fourth module is used to select key points from the block embedding using Fourier transform and inverse Fourier transform after input to the encoder, perform cross-attention operation on the block embedding and the key point embedding, perform self-attention operation on the block embedding, and then perform transformer module operation;

[0015] The fifth module is used to input the time series features extracted by the multi-layer transformer module into the linear layer for electromagnetic signal type recognition.

[0016] In a third aspect, the present invention provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method described in the first aspect when executing the program.

[0017] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in the first aspect.

[0018] In a fifth aspect, the present invention provides a computer program product, comprising a computer program, which implements the steps of the method described in the first aspect when executed by a processor.

[0019] Compared with the existing technology, the significant advantage of the present invention is that it proposes a method for constructing a time series basic model based on representation enhancement, thereby solving the problem of insufficient block representation of the time series basic model. This method makes up for the shortcomings of the sequence block-based method, such as insufficient block representation learning and low classification accuracy. By improving the block representation capability, the accuracy of time series classification is effectively improved. In addition, this method can achieve more accurate results in multiple practical application scenarios such as communications, radar, and medical care, and has strong practical value and broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1The overall flow chart of the method for building a time series basic model based on representation enhancement.

[0021] Figure 2 Network framework diagram of the method for building a temporal basic model based on representation enhancement.

[0022] Figure 3 Filter sub-flowgraphs for fine-grained information. DETAILED DESCRIPTION

[0023] Combine Figures 1 to 3 This paper proposes a method for constructing a time-series basic model based on representation enhancement and an electromagnetic recognition system. The method aims to optimize the block embedding representation of the time-series basic model, improve its ability to extract electromagnetic signal features, and thus enhance the recognition performance of electromagnetic signal types. The method specifically includes the following steps:

[0024] (1) Using electromagnetic signal data containing multi-dimensional time series features as the original sample, a time series basic model is constructed.

[0025] The input sample X is a time series data with multi-dimensional time series features, and its dimensions are Here, T represents the length of the time series, that is, the total number of time points; C represents the dimensionality of the variable, that is, the number of features observed at each time point. For example, in an electromagnetic signal monitoring scenario, X can be multidimensional data such as the amplitude and frequency of electromagnetic signals recorded by multiple sensors at different times. The output sample Y is the electromagnetic signal type label corresponding to the input sample X. Assuming there are D different electromagnetic signal types, then Y is a vector of length D, where each element represents the probability that the input sample X belongs to the corresponding electromagnetic signal type (one-hot encoding is typically used during training, that is, the element corresponding to the correct category is 1, and the rest are 0).

[0026] The model adopts a layered architecture, consisting of an input layer, a feature extraction layer, and a linear layer. The input layer receives multidimensional time series data X and performs preprocessing. The feature extraction layer contains a block embedding module, an enhancement module, and an encoder module. The block embedding module divides the input data into blocks and passes it through a linear mapping layer (involving a weight matrix W linear and the bias term b linear ) and one-dimensional convolution layer (involving convolution kernel weight W conv and the bias term b conv ) to implement the embedding operation. The enhancement module performs key point selection and nonlinear enhancement on the initial block embedding. The encoder module uses Fourier transform and inverse Fourier transform to select key points, and performs feature extraction through the cross attention layer (involving the query matrix Q, key matrix K and value matrix V) and the self-attention layer, and then further operates through the transformer module. The linear layer is based on the feature extraction layer (involving the weight matrix W pred and the bias term bpred ) performs electromagnetic signal recognition on the output of the feature extraction layer. Each layer is connected sequentially. Data flows into the input layer, is processed by the feature extraction layer, and then outputs the classification result T by the linear layer.

[0027] (2) The input time series is divided into multiple blocks, and each block is linearly mapped and one-dimensionally convolved to obtain the initial block embedding and point embedding.

[0028] First, the input multidimensional time series data Preprocessing is performed and linear interpolation is used to fill in missing values. For the missing value x of variable c at time point t t,c , and the interpolation result is:

[0029]

[0030] Among them, x (t-1),c and x (t+1),c are the observed values ​​of variable c at time points t-1 and t+1 respectively.

[0031] Next, the filled data is standardized using the formula:

[0032]

[0033] Among them, μ c and σ c are the mean and standard deviation of variable c respectively.

[0034] Then, the normalized time series data is divided into multiple fixed-length blocks in chronological order. Let the length of each block be P, then the number of blocks after division is The pth block X p It can be expressed as:

[0035]

[0036] Then, for each block X p Perform linear mapping and map it to high-dimensional space. The formula is:

[0037] z p,linear =W linear vec(X p )+b linear

[0038] Among them, d1 is the linear layer feature dimension, is a learnable linear mapping weight matrix, is the bias term, and vec(·) means flattening the matrix into a vector.

[0039] At the same time, for each block X pPerform a one-dimensional convolution operation with a convolution kernel size of k. The formula is:

[0040] z p,conv =Conv1D(X p ,W conv ,b conv )

[0041] Among them, d2 is the feature dimension of the convolution layer, is the convolution kernel weight, is the convolution bias, z p,conv It is the feature representation after convolution.

[0042] The linear mapping result z p,linear And the one-dimensional convolution result z p,conv Splice and get the initial block embedding z p :

[0043]

[0044] (3) Before inputting into the encoder, key points of the initial block embedding are selected and nonlinearly enhanced to obtain the initial enhanced block embedding.

[0045] First, embed z into the initial block p Calculate the significance score of each element. The significance score can be calculated using a gradient-based method or an activation-based method. For example, for the activation-based method, let z p,i is the block embedding z p The i-th element of i The calculation formula is:

[0046] S i =ReLU(W s z p,i +b s )

[0047] in, is a learnable weight vector, b s is the bias term.

[0048] Then, key points are selected based on the significance score. A threshold τ is set, and the index corresponding to the element with a significance score greater than τ is used as the key point index set S key .

[0049] Perform nonlinear enhancement on the selected key points, the formula is:

[0050]

[0051] Among them, α is a learnable parameter used to adjust the degree of enhancement to obtain the initial enhanced block embedding z p'.

[0052] (4) After inputting into the encoder, the block embedding is subjected to Fourier transform and inverse Fourier transform to select key points, the block embedding and key point embedding are subjected to cross attention operation, and then the block embedding is subjected to self-attention operation, and then the transformer module operation is performed.

[0053] The encoder receives the initial enhancement block embedding z p' First, the block embedding is processed through the Fourier transform layer. The Fourier transform formula is:

[0054]

[0055] where z p,n' is the block embedding z p' The nth element in F s is the complex value of the sth frequency component.

[0056] Then use the amplitude judgment layer, according to the amplitude |F s |Select key frequency points and set a frequency threshold τ f , the amplitude is greater than τ f The index corresponding to the frequency component is used as the key frequency point index set S f .

[0057] Then, the features corresponding to the key frequency points are inversely transformed through the inverse Fourier transform layer to obtain the key point embedding:

[0058] z key =IFFT({F s |s∈S f})

[0059] Where IFFT stands for Inverse Fourier Transform.

[0060] Then enter the cross attention layer, whose calculation formula is:

[0061]

[0062] Here Q is embedded in z by the block p' Generated, K and V are embedded in z by key points key Generate, d k is the dimension of the key.

[0063] Entering the self-attention layer again, the calculation formula is:

[0064]

[0065] where Q, K and V are all embedded in z by blocks p' generate.

[0066] Finally, it passes through the multi-layer transformer module, which includes operations such as multi-layer perceptron (MLP) and layer normalization, and outputs the encoded feature representation z encoder .

[0067] (5) The time series features extracted by the multi-layer transformer module are input into the linear layer for electromagnetic signal type recognition.

[0068] The linear layer receives the feature representation z output by the encoder encoder , to classify electromagnetic signal types. The formula of the linear layer is:

[0069]

[0070] in, is the learnable weight matrix, D is the total number of electromagnetic signal types, d encoder is the dimension of the encoder output feature representation; is the bias term; is a vector of length D, each element of which represents the unnormalized score (also called logits) of the input sample belonging to the corresponding electromagnetic signal type.

[0071] In order to convert the unnormalized scores into probability distributions, the softmax function is usually used, as shown in the following formula:

[0072]

[0073] in, represents the probability that the input sample belongs to the jth type of electromagnetic signal, is a vector The jth element of .

[0074] During the training process, the cross-entropy loss function is used as the loss function to measure the difference between the probability distribution predicted by the model and the actual category label. Assume that there are M time series samples, and for the i-th sample, its actual category label is y i (usually one-hot encoding is used, i.e. i is a vector of length D, where D is the total number of categories, the element corresponding to the correct category is 1, and the rest are 0). The formula for the cross entropy loss function is:

[0075]

[0076] Among them, y i,j is the actual category label vector y of the i-th sample i The jth element of It is the probability that the model predicts that the i-th sample belongs to the j-th type of electromagnetic signal. is the probability distribution vector obtained by the model's output for the i-th sample after processing it with the softmax function. By minimizing the cross-entropy loss function, the model can continuously adjust its parameters to make the predicted probability distribution as close as possible to the actual category label distribution, thereby improving the accuracy of electromagnetic signal type recognition.

[0077] Based on the same inventive concept, the present invention also proposes an electromagnetic recognition system based on a time-series basic model with enhanced representation, which is used to implement the above method and identify electromagnetic signals. The system includes:

[0078] The first module uses electromagnetic signal data containing multi-dimensional time series features as the original sample to build a time series basic model;

[0079] The second module divides the input time series into multiple blocks, performs linear mapping and one-dimensional convolution on each block, and obtains the initial block embedding and point embedding;

[0080] The third module selects key points of the initial block embedding and performs nonlinear enhancement on it before inputting it into the encoder to obtain the initial enhanced block embedding;

[0081] The fourth module, after inputting into the encoder, uses Fourier transform and inverse Fourier transform to select key points for the block embedding, performs cross attention operation on the block embedding and the key point embedding, and then performs self-attention operation on the block embedding, and then performs transformer module operation;

[0082] In the fifth module, the time series features extracted by the multi-layer transformer module are input into the linear layer for electromagnetic signal type recognition.

[0083] The specific implementation methods of the first to fifth modules are the same as the aforementioned method steps and will not be repeated here.

[0084] This paper addresses the issues in long-term electromagnetic signal processing related to time series, where block annotation learning ignores fine-grained information and block representation learning is insufficient, leading to inaccurate electromagnetic signal feature extraction and low type recognition accuracy. A method for constructing a time series base model and an electromagnetic recognition system based on representation enhancement is proposed. This method utilizes a key information screening mechanism to first identify key points within blocks of the electromagnetic signal time series and incorporate the fine-grained information contained in these key points into block features to construct an enhanced block representation. Subsequently, the time series base model is trained based on this enhanced block representation, and the recognition performance of the electromagnetic recognition system is optimized. This strategy effectively overcomes the accuracy issues that existing methods suffer from inadequate block representation in long-term electromagnetic signal processing and recognition tasks. It demonstrates excellent performance in long-term electromagnetic signal processing and recognition tasks, particularly in radar signal target recognition scenarios. It can also be further applied to a variety of practical tasks, such as communication signal analysis and medical physiological signal monitoring and recognition.

[0085] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for constructing a time series basic model based on representation enhancement, characterized in that: The steps include: Step 1: Using electromagnetic signal data containing multi-dimensional time series features as original samples, a time series basic model is constructed; Step 2: Divide the input time series into multiple blocks, perform linear mapping and one-dimensional convolution on each block, and obtain the initial block embedding and point embedding; Step 3: Before inputting to the encoder, key points of the initial block embedding are selected and nonlinearly enhanced to obtain the initial enhanced block embedding; Step 4: After inputting into the encoder, the block embedding is subjected to Fourier transform and inverse Fourier transform to select key points, the block embedding and key point embedding are subjected to cross attention operation, and then the block embedding is subjected to self-attention operation, and then the transformer module operation is performed; In step 5, the time series features extracted by the multi-layer transformer module are input into the linear layer for electromagnetic signal type recognition.

2. The method for constructing a time series basic model based on representation enhancement according to claim 1, characterized in that: In step 1, electromagnetic signal data containing multi-dimensional time series features is used as the original sample to construct a time series basic model, including the following steps: The input sample X is a time series data with multi-dimensional time series features, and its dimensions are T represents the length of the time series, that is, the total number of time points; C represents the dimensionality of the variable, that is, the number of features observed at each time point; the output sample Y is the electromagnetic signal type label corresponding to the input sample X. Assuming there are D different electromagnetic signal types, then Y is a vector of length D, where each element represents the probability that the input sample X belongs to the corresponding electromagnetic signal type; The model adopts a layered architecture and consists of an input layer, a feature extraction layer, and a linear layer. The input layer receives and preprocesses multidimensional time series data X. The feature extraction layer contains a block embedding module, an enhancement module, and an encoder module. The block embedding module blocks the input data and implements the embedding operation through a linear mapping layer and a one-dimensional convolutional layer. The enhancement module selects key points and performs nonlinear enhancement on the initial block embedding. The encoder module uses Fourier transform and inverse Fourier transform to select key points, extracts features through a cross-attention layer and a self-attention layer, and then performs further calculations through a transformer module. The linear layer performs electromagnetic signal recognition on the output of the feature extraction layer based on the feature extraction layer. The layers are connected in sequence, and data flows in from the input layer. After being processed by the feature extraction layer, the linear layer outputs the classification result Y.

3. The method for constructing a time series basic model based on representation enhancement according to claim 2, characterized in that: In step 2, the input time series is divided into multiple blocks, and each block is linearly mapped and one-dimensionally convolved to obtain the initial block embedding and point embedding, including the following steps: The input multidimensional time series data Preprocessing is performed and linear interpolation is used to fill in missing values; for the missing value x of variable c at time point t t,c , and the interpolation result is: Among them, x (t-1),c and x (t+1),c are the observed values ​​of variable c at time points t-1 and t+1 respectively; The filled data is standardized, and the formula is: Among them, μ c and σ c are the mean and standard deviation of variable c respectively; The standardized time series data is divided into multiple fixed-length blocks in chronological order; let the length of each block be P, then the number of blocks after division is The pth block X p Expressed as: Then, for each block X p Perform linear mapping and map it to high-dimensional space. The formula is: z p,linear =W linear ·vec(X p )+d linear in, is a learnable linear mapping weight matrix, d1 is the linear layer feature dimension, is the bias term, vec(·) means flattening the matrix into a vector; For each block X p Perform a one-dimensional convolution operation with a convolution kernel size of k. The formula is: z p,conv =Conv1D(X p ,W conv ,b conv ) in, is the convolution kernel weight, d2 is the convolution layer feature dimension, is the convolution bias, z p,conv It is the feature representation after convolution; The linear mapping result z p,linear And the one-dimensional convolution result z p,conv Splice and get the initial block embedding z p :

4. The method for constructing a time series basic model based on representation enhancement according to claim 3, characterized in that: In step 3, before inputting to the encoder, key points are selected for the initial block embedding and nonlinear enhancement is performed to obtain the initial enhanced block embedding, including the following steps: Embed z into the initial block p Calculate the significance score of each element; Select key points based on the significance score; set a threshold τ and take the index corresponding to the element with significance score greater than τ as the key point index set S key ; Perform nonlinear enhancement on the selected key points, the formula is: Among them, z p,i is the block embedding z p The i-th element of , α is a learnable parameter used to adjust the degree of enhancement to obtain the initial enhanced block embedding z p' .

5. The method for constructing a time series basic model based on representation enhancement according to claim 4, characterized in that: In step 4, after inputting to the encoder, the block embedding is subjected to Fourier transform and inverse Fourier transform to select key points, the block embedding and the key point embedding are subjected to cross attention operation, and then the block embedding is subjected to self-attention operation, and then the transformer module operation is performed, including the following steps: The encoder receives the initial enhancement block embedding z p' ; First, the block embedding is processed through the Fourier transform layer. The Fourier transform formula is: where z p,n' is the block embedding z p' The nth element in F s is the complex value of the sth frequency component; Use the amplitude to judge the layer, according to the amplitude |F s |Select key frequency points and set a frequency threshold τ f , the amplitude is greater than τ f The index corresponding to the frequency component is used as the key frequency point index set S f ; Through the inverse Fourier transform layer, the features corresponding to the key frequency points are inverse Fourier transformed to obtain the key point embedding: z key =IFFT({F s ∣s∈S f }) Where IFFT stands for inverse Fourier transform; Then enter the cross attention layer, whose calculation formula is: Here Q is embedded in z by the block p' Generated, K and V are embedded in z by key points key Generate, d k is the dimension of the key; Entering the self-attention layer again, the calculation formula is: where Q, K and V are all embedded in z by blocks p' generate; Finally, after passing through the multi-layer transformer module, which contains a multi-layer perceptron and layer normalization operation, the encoded feature representation z is output. encoder .

6. The method for constructing a time series basic model based on representation enhancement according to claim 5, characterized in that: In step 5, the time series features extracted by the multi-layer transformer module are input into the linear layer to perform electromagnetic signal type recognition, which includes the following steps: The linear layer receives the feature representation z output by the encoder encoder , to classify electromagnetic signal types; the formula of the linear layer is: in, is the learnable weight matrix, D is the total number of electromagnetic signal types, d encoder is the dimension of the encoder output feature representation; is the bias term; Is a vector of length D, each element of which represents the unnormalized score of the input sample belonging to the corresponding electromagnetic signal type; Use the softmax function to convert the unnormalized scores into probability distributions. The formula is as follows: in, represents the probability that the input sample belongs to the jth type of electromagnetic signal, is a vector The jth element of ; During the training process, the cross-entropy loss function is used as the loss function to measure the difference between the probability distribution predicted by the model and the actual category label; assuming there are M time series samples, for the i-th sample, its actual category label is y i , using one-hot encoding, y i Is a vector of length D, where D is the total number of categories, the element corresponding to the correct category is 1, and the rest are 0; the formula for the cross entropy loss function is: Among them, y i,j is the actual category label vector y of the i-th sample i The jth element of is the probability that the model predicts that the i-th sample belongs to the j-th type of electromagnetic signal; It is the probability distribution vector obtained by the model's output for the i-th sample after being processed by the softmax function.

7. A time series based model electromagnetic signal recognition system based on representation enhancement, characterized in that: For implementing the method according to any one of claims 1 to 6, the system comprises: The first module is used to construct a time series basic model using electromagnetic signal data containing multi-dimensional time series features as original samples; The second module is used to divide the input time series into multiple blocks, perform linear mapping and one-dimensional convolution on each block, and obtain the initial block embedding and point embedding; The third module is used to select key points of the initial block embedding and perform nonlinear enhancement on it before inputting it into the encoder to obtain the initial enhanced block embedding; The fourth module is used to select key points from the block embedding using Fourier transform and inverse Fourier transform after input to the encoder, perform cross-attention operation on the block embedding and the key point embedding, perform self-attention operation on the block embedding, and then perform transformer module operation; The fifth module is used to input the time series features extracted by the multi-layer transformer module into the linear layer for electromagnetic signal type recognition.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images