A radar time series data target identification method based on an improved ShapeFormer architecture

Abstract

The application provides a radar time series data target identification method based on an improved ShapeFormer architecture, and the feature extraction method combining Shapelet discovery, differential embedding and a double-channel Transformer is proposed based on the improved ShapeFormer architecture, so that the local discriminability and the global dependency are considered. Meanwhile, a complete preprocessing process including missing value filling, standardization and T-SMOTE time series oversampling is constructed, so that the size of the minority class sample is effectively expanded. In combination with a cross-entropy loss and a sample weight optimization strategy, the learning ability of the model under the condition of class imbalance is improved. The application has reliable design, wide prospect and great application space in the field of radar target identification.

Description

Technical Field

[0001] This invention relates to the field of radar target recognition technology, specifically to a radar time-series data target recognition method based on an improved ShapeFormer architecture. Background Technology

[0002] Radar signal data is widely used in typical application scenarios such as meteorological monitoring and traffic surveillance, and target recognition methods based on radar signal data have significant application value. However, traditional radar signal processing methods mainly rely on rule-based algorithms, which lack flexibility and have poor recognition performance when faced with various types of complex targets.

[0003] In recent years, deep learning methods, especially Transformer-based models, have been successfully applied to the analysis of time series data. Existing methods primarily focus on extracting global features and simple spatiotemporal patterns, neglecting the unique and discriminative details specific to target categories. In radar signals, due to differences in target shape and trajectory, the temporal characteristics of targets exhibit significant category-specific differences. Therefore, traditional methods cannot fully capture category-specific details, resulting in low recognition accuracy in complex environments. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention proposes a radar time-series data target identification method based on an improved ShapeFormer architecture, comprising the following steps:

[0005] Step 1: Obtain the original radar time series dataset;

[0006] Step 2: Preprocess the original radar time series dataset and divide the preprocessed radar time series dataset into training set, validation set and test set;

[0007] Step 3: Perform Shapelet discovery on the training set to find local subsequences with class discriminative power and store them in the Shapelet pool;

[0008] Step 4: Using the input sequence Shapelet filtering is performed on candidate subsequences in the Shapelet pool to generate the best-matching subsequence and the difference vector between the candidate subsequence and the best-matching subsequence; during training, the input sequence... The training set consists of radar time-series data; the input sequence is used during actual identification. Radar time-series data for target identification;

[0009] Step 5: Add position encoding to the difference vector, wherein the position encoding includes the start position, end position, and variable index of the candidate subsequence corresponding to the difference vector;

[0010] Step 6: Input the difference vector after the superimposed position encoding process into the Transformer encoder with class-specific Transformer module to extract the class-specific token;

[0011] Step 7: Process the input sequence Perform general Transformer processing to obtain a general token; during training, the input sequence... The training set consists of radar time-series data; the input sequence is used during actual identification. Radar time-series data for target identification;

[0012] Step 8: Obtain the target recognition result based on the class-specific token obtained in Step 6 and the general token obtained in Step 7; during training, use the target recognition result and the corresponding labels of the time series in the training set to perform cross-entropy training, and verify the trained model through the validation set, and evaluate the indicator through the test set to obtain a target recognition model that meets the evaluation requirements.

[0013] Furthermore, the data preprocessing in step 2 includes missing value imputation, standardization, and data oversampling.

[0014] Furthermore, the data oversampling process is as follows:

[0015] Suppose the radar time series data after missing value imputation and standardization is: Each sample is a multidimensional time series of length T:

[0016]

[0017] The feature vector at each time point j is ;

[0018] Set the lead time to Then, for each preceding time, we define as: Positive samples and It is a temporal neighborhood; for a temporal neighborhood pair New samples are synthesized using linear interpolation:

[0019]

[0020] in These are the interpolation coefficients.

[0021] Furthermore, in step 3, Shapelet discovery is performed on the data in the training set to obtain a set of candidate subsequences, where similarity is calculated using a time series distance metric.

[0022] Furthermore, in step 4, the input sequence is used. The process of performing Shapelet filtering on candidate subsequences in the Shapelet pool and generating difference vectors is as follows:

[0023] Input sequence With candidate subsequences stored in the Shapelet pool The index of the most matching subsequence is fed into the Shapelet filter. ;

[0024] Will its best matching subsequence The difference vector is obtained by mapping to an embedding space of the same dimension through a linear layer and then subtracting the layers. .

[0025] Furthermore, in step 5, the process of adding positional encoding to the difference vector is as follows: the start position, end position, and variable index of the candidate subsequence are obtained by one-hot encoding and linear processing respectively, and then the three positional encodings are superimposed on the difference vector.

[0026] Furthermore, in step 6, the specific process of inputting the differential vector after superimposed position encoding into the Transformer encoder with class-specific Transformer modules to extract class-specific tokens is as follows:

[0027] According to the formula

[0028]

[0029]

[0030] Calculate the output of the i-th token ,in , , , are linear projection matrices of q, k, and v, respectively; For class-specific embedding dimensions; These are the attention weights after softmax. For the j-th difference vector, Let be the i-th difference vector.

[0031] Furthermore, in step 7, the input sequence... The specific process of obtaining a general token through general Transformer processing is as follows:

[0032] Input sequence The input is fed into a general Transformer, and after being concatenated by two ConvBlocks, the serialized features are obtained. Serialization features Add learnable position embedding Then it is input into the multi-head attention mechanism to obtain ,right A general token is obtained by performing average pooling.

[0033] Furthermore, in step 7, the batch normalization function BatchNorm and the activation function used for a single ConvBlock = Conv1D → BatchNorm → GELU are... They are shown below:

[0034]

[0035]

[0036] In the formula, It is the input sequence. and These are the inputs The mean and standard deviation, It is a constant.

[0037] Furthermore, in step 8, the tokens output by the class-specific Transformer and the general Transformer are concatenated and then passed through a linear layer + softmax to obtain the target recognition result.

[0038] Beneficial effects

[0039] This invention provides an efficient and high-precision radar target recognition method to address the problems of imbalanced target samples and missing data when performing radar target recognition tasks. Based on an improved ShapeFormer architecture, this method proposes a feature extraction approach combining "Shapelet discovery + differential embedding + dual-channel Transformer," balancing local discriminability and global dependency. Simultaneously, a complete preprocessing workflow is constructed, including missing value imputation, standardization, and T-SMOTE time series oversampling, effectively expanding the minority class sample size. Furthermore, by combining cross-entropy loss and sample weight optimization strategies, the model's learning ability under imbalanced conditions is enhanced. This invention is reliable, has broad prospects, and possesses significant application potential in the field of radar target recognition.

[0040] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0041] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0042] Figure 1 Schematic diagram of the data preprocessing module;

[0043] Figure 2 : Schematic diagram of a Shapelet filter;

[0044] Figure 3 : A diagram illustrating the class-specific Transformer and the general Transformer framework;

[0045] Figure 4 : Improved Shapeformer network framework diagram. Detailed Implementation

[0046] The embodiments of the present invention are described in detail below. These embodiments are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0047] This embodiment provides a target identification method for radar time-series data based on an improved Shapeformer architecture, used for effective target identification of radar time-series data in the presence of missing data and imbalanced data categories. The specific steps are as follows:

[0048] Step 1: Obtain the dataset:

[0049] The original radar time series dataset is obtained by actually collecting radar time series data or by acquiring raw radar time series data from an open-source dataset platform:

[0050]

[0051] in Indicates the first A time series of samples, with length T and dimension d. For the first The target identification label corresponding to each time series.

[0052] Step 2: Preprocess the original radar time series dataset:

[0053] The steps for preprocessing the data in the original radar time series dataset include:

[0054] Step 2.1: Imputing missing values:

[0055] For time-series radar signal data, assuming that in the first... Time series In, a certain feature dimension The t-th time point is missing, i.e. Then the linear interpolation formula is:

[0056]

[0057] In the formula Indicates the first The first in the time series The moment of the first Each feature value corresponds to a dataset matrix. One of the elements, where This indicates the time index of the currently missing value. Indicates the feature index dimension. This represents the index of the two most recent non-missing time points before and after the missing value. and In the i-th time series, the i-th time series is... These features at a given time point The known value.

[0058] Step 2.2: Standardization Processing:

[0059] To ensure scale consistency among features, each feature in the input radar time series data is standardized to ensure that the mean of each feature is 0 and the variance is 1. The formula for data table transformation is as follows:

[0060]

[0061] In the formula This represents the converted numerical value. This represents the mean of feature k across the entire dataset. This represents the standard deviation of characteristic k.

[0062] Step 2.3: Data oversampling:

[0063] T-SMOTE (Temporal-oriented Synthetic Minority Over-sampling Technique) is a method that extends the classic SMOTE idea (interpolating between minority class samples to generate new synthetic samples) to time series / temporal samples. Its core idea is to interpolate minority class samples along the temporal neighborhood in the temporal dimension, thereby generating new temporal samples to alleviate class imbalance, enhance the model's ability to discriminate the minority class, and simultaneously maintain temporal continuity and the structure of temporal features as much as possible.

[0064] Here, the radar time series data processed in steps 2.1 and 2.2 are represented as a minority class sample set. Each sample is a multidimensional time series of length T:

[0065]

[0066] The feature vector at each time point j is .

[0067] Set the leading time to Define each leader time as Positive samples and It is a temporal neighborhood.

[0068] For a time neighborhood pair T-SMOTE uses linear interpolation to synthesize more samples:

[0069]

[0070] in These are the interpolation coefficients.

[0071] The addition and scalar multiplication here are performed element-wise, point-by-point and dimension-by-dimensional, meaning that for each time step t:

[0072]

[0073] in The feature dimension representing a specific point in time is eigenvectors, .

[0074] Geometric meaning: When using the same scalar When interpolating the entire sequence, the generated sequence is located in two time neighborhoods. The convex combination of the connecting lines (i.e., points on the line segments) helps to maintain the temporal consistency and local smoothness of the entire sequence.

[0075] Step 2.4: After the above steps, the radar time series dataset is divided into a training set, a validation set, and a test set.

[0076] Step 3: Perform Shapelet discovery using the radar time series in the training set, find local subsequences with class discriminative power, and store them in the Shapelet pool.

[0077] The training set data after step 2 preprocessing The input is fed into the Shapelet to search for a set of candidate subsequences. S = [ S 1 , S 2 , … , S l ] ,in The subsequences are then stored in a Shapelet pool, where similarity is calculated using a time-series distance metric, as shown in the following formula:

[0078] P S D ( X , S ) = m i n j = 1 T − l + 1 C I D ( X [ j : j + l − 1 ] , S )

[0079] In the formula It is a commonly used time series distance metric to measure the similarity of subsequences and to find similarities with Shapelets. The closest subsequence, using Ensure that it is insensitive to differences in complexity. X [ j : j + l − 1 ] Indicates index The starting length is subsequence.

[0080] Step 4: Using the input sequence Shapelet filtering is performed on the candidate subsequences in the Shapelet pool, and a difference vector is generated:

[0081] Input sequence With candidate subsequences stored in the Shapelet pool The index of the most matching subsequence is fed into the Shapelet filter:

[0082] i n d e x = a r g m i n j = 0 T − l + 1 C I D ( X [ j : j + l ] , S i )

[0083] I i = X [ i n d e x : i n d e x + l ]

[0084] in In the input sequence Chinese envoy Minimum starting index, The corresponding best-matching subsequence, also of length . .

[0085] Will its best matching subsequence The difference vector is obtained by mapping to an embedding space of the same dimension through a linear layer and then subtracting the layers. As a token:

[0086]

[0087] in and For a linear projector, subsequences (dimensions) or ×1) Mapped to a length of The vector space. Used for differential embedding.

[0088] During training, the time series data from the training set are used as the input sequence. When actually performing recognition, the input sequence The radar time series data is for target identification.

[0089] Step 5: For the difference vector Add position encoding:

[0090] Due to the difference vector Since it does not contain location information itself, the three types of location information are encoded and added together. Position encoding is performed on the above, and the three types of position information are: starting position End position as well as The index of the variable :

[0091]

[0092]

[0093] This indicates that the position (integer index) is encoded using one-hot encoding. This means that the positional encoding vector is obtained by passing one-hot encoding through a linear layer, and a set of learnable embeddings is made for the start position, end position, and variable index, and then superimposed on the vector. superior.

[0094] Step 6: Class-specific Transformer coding:

[0095] After overlay position encoding processing U = [ U 1 , U 2 , … , U g ] The class-specific Shapelet representation is extracted by feeding the data into a Transformer encoder with a class-specific Transformer module (multi-head attention), and the multi-head attention forms the output. Z spe = [ Z 1 spe , . . . , Z g spe ] The calculation form for a single token is as follows:

[0096]

[0097]

[0098] in , , , are linear projection matrices of q, k, and v, respectively; For class-specific embedding dimensions; These are the attention weights after softmax. This is the output of the i-th token after encoding.

[0099] Step 7: General Transformer processing:

[0100] Input sequence The input is fed into a general Transformer, and then passed through two concatenated ConvBlocks (each ConvBlock = Conv1D → BatchNorm → GELU). The kernel dimension of the first Conv1D is... The second Conv1D convolution kernel dimension is Then the obtained serialized features Add learnable position embedding The input is obtained into the multi-head attention mechanism. Finally, average pooling is performed to obtain a universal token.

[0101]

[0102]

[0103]

[0104]

[0105] in The width of the convolution kernel; The output sequence of ConvBlock (each position is) (Wei's token); For learnable location embeddings (and) (Length matching) The generic token (vector) obtained through average pooling is used for subsequent concatenation.

[0106] The batch normalization function BatchNorm and activation function used in the above formula They are shown below:

[0107]

[0108]

[0109] In the formula, It is the input sequence. and These are the inputs The mean and standard deviation, It is a constant.

[0110] Similarly, during training, the time series data from the training set are used as the input sequence. When actually performing recognition, the input sequence The radar time series data is for target identification.

[0111] Step 8: Concatenate the tokens output by the class-specific Transformer and the general Transformer, and then pass them through a linear layer + softmax to obtain the target recognition result:

[0112]

[0113]

[0114] In the formula A class-specific final token; The final token is the general one; Linear is the category header (fully connected) mapped to the number of categories dimension.

[0115] During training, cross-entropy training is performed based on the labels corresponding to the time series data in the training set. The cross-entropy loss function is:

[0116]

[0117] Using standard cross-entropy loss, we obtain the training model with the minimum loss. And the training model was tested using a validation set. Perform validation to determine if early stopping is necessary. After training is complete, feed the time series data from the test set into the training model. The trained model is evaluated using metrics such as accuracy, recall, and precision. The final result is a target recognition model that meets the overall evaluation requirements. (Recall is mentioned in the original text, but its relevance to the main point is unclear.) and accuracy They are shown below:

[0118]

[0119] In the formula, It is the number of samples that the model correctly predicts as positive. It is the number of samples that the model incorrectly predicted as positive. It represents the number of samples that the model incorrectly predicted as negative.

[0120] The actual radar time series data to be predicted is used as the input sequence. Input the target recognition model that meets the comprehensive evaluation requirements to obtain the target recognition result.

[0121] Based on the above method, this embodiment also provides a radar time series data target recognition system with an improved Shapeformer architecture, including a data preprocessing module, a Shapelet discovery module and a target recognition module;

[0122] The data preprocessing module retrieves and preprocesses the original radar time series data, treats the radar signal as multivariate time series data, and obtains preprocessed radar time series data; and divides the preprocessed radar time series data into training set, validation set and test set.

[0123] The Shapelet discovery module takes the training set as input, discovers local subsequences with class discriminative power through Shapelet discovery, and stores them in the Shapelet pool;

[0124] The target recognition module includes a class-specific Transformer module and a general Transformer module; the class-specific Transformer module processes the input sequence. The input sequence is processed with candidate subsequences in the Shapelet pool to obtain a class-specific token; the general Transformer module processes the input sequence. The process is performed to obtain a general token; the class-specific token and the general token are used to perform target recognition and prediction to obtain the prediction result.

[0125] During training, the input sequence of the class-specific Transformer module... For radar time-series data in the training set, the input sequence of the class-specific Transformer module is used during actual target recognition. The radar time series data to be identified.

[0126] This invention proposes a radar time-series data target recognition method based on an improved Shapeformer architecture. It introduces an improved Shapeformer architecture into the traditional Transformer framework, utilizing a Shapelet discovery and filtering mechanism to first extract local sub-sequence features with class discriminative power. Then, it performs fine-grained feature modeling through a class-specific Transformer encoder and combines this with global temporal features extracted by a general Transformer to achieve a dual representation of "local differences + global dependencies." This structure significantly improves the ability to capture complex temporal features in radar target recognition, compensates for the shortcomings of traditional methods in fine-grained difference recognition, and thus effectively improves recognition accuracy and stability.

[0127] To address the common class imbalance problem in radar time-series data, this invention employs T-SMOTE time-series oversampling technology combined with the Shapeformer architecture. T-SMOTE generates high-fidelity minority class samples within the temporal neighborhood, significantly expanding the diversity of training data distribution, and complements the local-global feature extraction mechanism of Shapeformer. T-SMOTE enhances the model's representation ability of minority class samples, while the improved Shapeformer framework, while enhancing sample diversity, can fully utilize its discriminative detail features, thereby achieving balanced recognition of all target types. This invention combines T-SMOTE with the improved Shapeformer framework to effectively improve target recognition performance in imbalanced radar time-series data.

[0128] Furthermore, this invention leverages the unique characteristics of radar time-series data (such as high dimensionality, strong noise interference, and significant differences in motion trajectories between target categories) by introducing differential embedding and multi-scale convolutional feature extraction mechanisms into the Shapeformer architecture. This enables the model to simultaneously focus on short-term local morphological changes and long-term global trends. Compared to traditional methods, this design is more adaptable to the non-stationarity and complex interference present in radar echo signals, significantly improving robustness in real-world environments and the ability to distinguish between different target categories.

[0129] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention without departing from the principles and spirit of the present invention.

Claims

1. A target recognition method for radar time-series data based on an improved ShapeFormer architecture, characterized in that: Includes the following steps: Step 1: Obtain the original radar time series dataset; Step 2: Preprocess the original radar time series dataset and divide the preprocessed radar time series dataset into training set, validation set and test set; Step 3: Perform Shapelet discovery on the training set to find local subsequences with class discriminative power and store them in the Shapelet pool; Step 4: Using the input sequence Shapelet filtering is performed on candidate subsequences in the Shapelet pool to generate the best-matching subsequence and the difference vector between the candidate subsequence and the best-matching subsequence; during training, the input sequence... The training set consists of radar time-series data; the input sequence is used during actual identification. Radar time-series data for target identification; Step 5: Add position encoding to the difference vector, wherein the position encoding includes the start position, end position, and variable index of the candidate subsequence corresponding to the difference vector; Step 6: Input the difference vector after the superimposed position encoding process into the Transformer encoder with class-specific Transformer module to extract the class-specific token; Step 7: Process the input sequence Perform general Transformer processing to obtain a general token; during training, the input sequence... The training set consists of radar time-series data; the input sequence is used during actual identification. Radar time-series data for target identification; Step 8: Obtain the target recognition result based on the class-specific token obtained in Step 6 and the general token obtained in Step 7; during training, use the target recognition result and the corresponding labels of the time series in the training set to perform cross-entropy training, and verify the trained model through the validation set, and evaluate the indicator through the test set to obtain a target recognition model that meets the evaluation requirements.

2. The radar time-series data target recognition method based on the improved ShapeFormer architecture according to claim 1, characterized in that: The data preprocessing in step 2 includes missing value imputation, standardization, and data oversampling.

3. The radar time-series data target recognition method based on the improved ShapeFormer architecture according to claim 2, characterized in that: The data oversampling process is as follows: Suppose the radar time series data after missing value imputation and standardization is: Each sample is a multidimensional time series of length T: The feature vector at each time point j is ; Set the lead time to Then, for each preceding time, we define as: Positive samples and It is a temporal neighborhood; for a temporal neighborhood pair New samples are synthesized using linear interpolation: in These are the interpolation coefficients.

4. The radar time-series data target recognition method based on the improved ShapeFormer architecture according to claim 1, characterized in that: In step 3, Shapelet discovery is performed on the data in the training set to obtain a set of candidate subsequences, where similarity is calculated using a time series distance metric.

5. The radar time-series data target recognition method based on the improved ShapeFormer architecture according to claim 1, characterized in that: In step 4, the input sequence is used. The process of performing Shapelet filtering on candidate subsequences in the Shapelet pool and generating difference vectors is as follows: Input sequence With candidate subsequences stored in the Shapelet pool The index of the most matching subsequence is fed into the Shapelet filter. ; Will its best matching subsequence The difference vector is obtained by mapping to an embedding space of the same dimension through a linear layer and then subtracting the layers. .

6. The radar time-series data target recognition method based on the improved ShapeFormer architecture according to claim 1, characterized in that: In step 5, the process of adding positional encoding to the difference vector is as follows: the start position, end position, and variable index of the candidate subsequence are obtained by one-hot encoding and linear processing respectively, and then the three positional encodings are superimposed on the difference vector.

7. The radar time-series data target recognition method based on the improved ShapeFormer architecture according to claim 1, characterized in that: In step 6, the specific process of inputting the differential vector after overlay position encoding into the Transformer encoder with class-specific Transformer module to extract the class-specific token is as follows: According to the formula Calculate the output of the i-th token ,in , , , are linear projection matrices of q, k, and v, respectively; For class-specific embedding dimensions; These are the attention weights after softmax. For the j-th difference vector, Let be the i-th difference vector.

8. The radar time-series data target recognition method based on the improved ShapeFormer architecture according to claim 1, characterized in that: In step 7, the input sequence is... The specific process of obtaining a general token through general Transformer processing is as follows: Input sequence The input is fed into a general Transformer, and after being concatenated by two ConvBlocks, the serialized features are obtained. Serialization features Add learnable position embedding Then it is input into the multi-head attention mechanism to obtain ,right A general token is obtained by performing average pooling.

9. The radar time-series data target recognition method based on the improved ShapeFormer architecture according to claim 8, characterized in that: In step 7, a single ConvBlock = Conv1D → BatchNorm → GELU uses the batch normalization function BatchNorm and the activation function. They are shown below: In the formula, It is the input sequence. and These are the inputs The mean and standard deviation, It is a constant.

10. The radar time-series data target recognition method based on the improved ShapeFormer architecture according to claim 1, characterized in that: In step 8, the tokens output by the class-specific Transformer and the general Transformer are concatenated and then passed through a linear layer + softmax to obtain the target recognition result.

Images