HRRP sequence identification method and device based on lightweight Transform and full convolutional network

Abstract

The invention provides an HRRP sequence identification method and device based on a lightweight Transform and a full convolutional network, and relates to the technical field of radar target automatic identification. The method comprises the following steps: inputting an HRRP sequence into a time feature encoder module to extract time features; inputting the HRRP sequence into a spatial feature full convolution module to extract spatial features; and inputting the time features and the space features into a decision fusion identification module of the Trans-FCN model to obtain a radar target identification result. According to the method, the time feature encoder module and the spatial feature convolution module are used for respectively extracting the correlation features of the time and the space of the HRRP target sequence, and meanwhile, the Transform is subjected to lightweight improvement, so that the model has relatively low calculation complexity and model parameters while keeping relatively high accuracy, the deployment of edge equipment is facilitated, and the method is suitable for large-scale popularization and application. And the identification performance on the HRRP data set with unbalanced categories is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of automatic radar target recognition, and in particular to a HRRP sequence recognition method and device based on a lightweight Transformer and a full convolutional network. Background Art

[0002] Radar target automatic recognition technology is an important radar signal processing technology. By analyzing the radar echo signal, the type of target can be automatically identified. This technology is widely used in both military and civilian fields. Radar high-resolution range profile is the vector sum of the target scattered echo obtained based on the radar broadband signal projected in the direction of the radar ray. It contains discernible information such as target structure and target size, so it is widely used in radar target recognition. When detecting moving targets, the radar and the target will move relative to each other, so the echo information of multiple continuous azimuth angles can be obtained, and a continuous HRRP (High Resolution Range Profile) sequence can be constructed. By modeling the correlation of the HRRP sequence, the dynamic characteristics of the target can be effectively extracted to achieve efficient target recognition tasks.

[0003] Traditional HRRP target recognition methods usually require manual selection and design of features, which is not only labor-intensive but also easily affected by personal experience and subjective judgment. In addition, traditional methods can usually only handle static, pre-set target categories, and their recognition performance is usually poor for dynamically changing, unknown target categories. With the continuous development of deep learning, methods based on convolutional neural networks and recurrent neural networks are also widely used for HRRP sequence recognition. However, current methods for HRRP sequence recognition are difficult to extract and fully utilize the global spatiotemporal features of HRRP sequences. Methods based on CNN (Convolutional Neural Networks) are limited by the size of the convolution kernel and have limitations in extracting global features of long sequences; methods based on RNN (Recurrent Neural Network) lose important features of the target due to the loss of memory state when extracting long sequence information. With the introduction of Transformer, the global correlation of the sequence is modeled using the attention mechanism, which can effectively extract the global information of the sequence, but the current methods lack lightweight considerations, and the number of parameters required for the model is huge, making it difficult to deploy on edge devices.

[0004] Domestic and foreign research on radar target recognition based on HRRP mainly focuses on two aspects, namely, the recognition of cooperative targets and the recognition of non-cooperative targets. The data of cooperative targets is relatively easy to obtain, and the amount of data is sufficient, while the data of non-cooperative targets is often difficult to obtain and is affected by the detection distance, detection environment and target attributes. In actual scenarios in the fields of national defense, military, aerospace, etc., the targets to be identified are often non-cooperative targets, and it is difficult to obtain a large number of samples and high signal-to-noise ratio data. The data set presents the characteristics of many sample types and unbalanced categories. Summary of the invention

[0005] In order to solve the technical problems that the existing technology lacks lightweight considerations, the number of parameters required for the model is huge, it is difficult to deploy on edge devices, the targets to be identified are often non-cooperative targets, it is difficult to obtain a large number of samples and high signal-to-noise ratio data, and the data set presents the characteristics of many sample types and imbalanced categories, the embodiment of the present invention provides a HRRP sequence recognition method and device based on a lightweight Transformer and a fully convolutional network. The technical solution is as follows:

[0006] On the one hand, a HRRP sequence recognition method based on a lightweight Transformer and a fully convolutional network is provided, the method is implemented by a HRRP sequence recognition device, and the method includes:

[0007] S1. Obtain a HRRP sequence to be identified.

[0008] S2. Input the HRRP sequence into the temporal feature encoder module of the Trans-FCN model to extract the temporal features of the HRRP sequence.

[0009] S3. Input the HRRP sequence into the spatial feature full convolution module of the Trans-FCN model to extract the spatial features of the HRRP sequence.

[0010] S4. Input the temporal features of the HRRP sequence and the spatial features of the HRRP sequence into the decision fusion recognition module of the Trans-FCN model to obtain the radar target recognition result.

[0011] Among them, the temporal feature encoder module in S2 inputs the HRRP sequence into the Trans-FCN model to extract the temporal features of the HRRP sequence, including:

[0012] A high-dimensional linear mapping is applied to the HRRP sequence to obtain a high-dimensional sequence; learnable position information is added to the high-dimensional sequence to obtain the input data of the temporal feature encoder; the output of the temporal feature encoder is obtained according to the input data of the temporal feature encoder and the temporal feature encoder; the output of the temporal feature encoder is input into the global average pooling layer of the temporal feature encoder module to obtain the temporal features of the HRRP sequence; wherein the temporal feature encoder includes a grouped linear transformation layer, a self-attention layer, a Dropout regularization layer, a first residual connection layer and a normalization layer, a lightweight feedforward neural network layer, and a second residual connection layer and a normalization layer.

[0013] The spatial feature full convolution module includes a first convolution block, a first squeeze excitation layer, a second convolution block, a second squeeze excitation layer, a third convolution block, and a pooling layer of the spatial feature full convolution module.

[0014] Optionally, obtaining a time feature encoder output according to the time feature encoder input data and the time feature encoder includes:

[0015] S21. Input the temporal feature encoder input data into the grouped linear transformation layer to obtain the self-attention layer input data.

[0016] S22. Input the self-attention layer input data into the self-attention layer to obtain the attention feature.

[0017] S23. Input the attention feature into the Dropout regularization layer, the first residual connection layer and the normalization layer to obtain the first residual connection layer output time feature.

[0018] S24. Input the output time feature of the first residual connection layer into the lightweight feedforward neural network layer to obtain the output time feature of the lightweight feedforward neural network layer.

[0019] S25. Input the temporal features output by the lightweight feedforward neural network layer into the second residual connection layer and the normalization layer to obtain the temporal feature encoder output.

[0020] Optionally, in S21, inputting the temporal feature encoder input data into the grouped linear transformation layer to obtain the self-attention layer input data includes:

[0021] The input data of the temporal feature encoder is grouped to obtain 4 groups of data, each of the 4 groups of data is dimensionally upgraded to obtain 4 groups of high-dimensional features, and the 4 groups of high-dimensional features are merged to obtain a high-dimensional vector.

[0022] The high-dimensional vectors are grouped to obtain 4 groups of data. Each of the 4 groups of data is reduced in dimension to obtain 4 groups of low-dimensional features. The 4 groups of low-dimensional features are concatenated to obtain the input data of the self-attention layer.

[0023] Optionally, the attention feature in S22 is as shown in the following formula (1):

[0024] (1)

[0025] In the formula, Represents the attention feature, represents the query vector, Represents the input data of the self-attention layer, represents the key vector, represents a value vector, , , represents the learnable weight matrix.

[0026] Optionally, in S3, the HRRP sequence is input into the spatial feature full convolution module of the Trans-FCN model to extract the spatial features of the HRRP sequence, including:

[0027] S31. Input the HRRP sequence into the first convolution block to obtain the output spatial features of the first convolution block.

[0028] S32. Input the spatial features output by the first convolution block into the first squeeze excitation layer to obtain the first squeeze excitation layer output.

[0029] S33. Input the output of the first squeeze-excitation layer into the second convolution block to obtain the output spatial features of the second convolution block.

[0030] S34. Input the spatial features output by the second convolution block into the second squeeze excitation layer to obtain the second squeeze excitation layer output.

[0031] S35. Input the output of the second squeeze-excitation layer into the third convolution block to obtain the output spatial features of the third convolution block.

[0032] S36. Input the spatial features output by the third convolution block into the pooling layer of the spatial feature full convolution module to obtain the spatial features of the HRRP sequence.

[0033] Among them, the first convolution block, the second convolution block and the third convolution block respectively include a one-dimensional convolution layer, a batch normalization layer and a ReLU activation function.

[0034] The size of the convolution kernel of the first convolution block is 8, the size of the convolution kernel of the second convolution block is 5, and the size of the convolution kernel of the third convolution block is 3.

[0035] Optionally, in S4, the temporal features of the HRRP sequence and the spatial features of the HRRP sequence are input into a decision fusion recognition module of the Trans-FCN model to obtain a radar target recognition result, including:

[0036] S41. Input the temporal features of the HRRP sequence and the spatial features of the HRRP sequence into the fully connected layer respectively, and classify them through the softmax function to obtain the temporal branch classification result and the spatial branch classification result.

[0037] S42, establish a learnable time branch decision weight matrix and a space branch decision weight matrix, and obtain a radar target recognition result according to the time branch classification result, the space branch classification result, the time branch decision weight matrix and the space branch decision weight matrix, as shown in the following formula (2):

[0038] (2)

[0039] In the formula, Indicates the radar target recognition result, represents the time branch decision weight matrix, represents the time branch classification result, represents the spatial branch decision weight matrix, Represents the spatial branch classification result.

[0040] Optionally, the process of constructing the loss function of the Trans-FCN model includes:

[0041] Combining the label smoothing strategy with the focal loss function, we get the combined loss function, as shown in the following formula (3):

[0042] (3)

[0043] In the formula, represents the combined loss function, Indicates the number of categories of the true label, represents the smoothing factor, Indicates The one-hot encoded labels of the classes, represents the balance factor, Represents the model for The predicted probability of the class, Represents the adjustment factor.

[0044] The balance factor and adjustment factor in the combined loss function are dynamically adaptively processed to obtain an adaptive focus loss function combined with a label smoothing strategy, as shown in the following formula (4):

[0045] (4)

[0046] In the formula, represents the adaptive focal loss function combined with the label smoothing strategy, represents the total number of samples of all categories, Indicates The number of samples of the class, represents the initial value of the dynamic adjustment factor, represents an adjustable hyperparameter.

[0047] On the other hand, a HRRP sequence recognition device based on a lightweight Transformer and a full convolutional network is provided, and the device is applied to a HRRP sequence recognition method based on a lightweight Transformer and a full convolutional network, and the device includes:

[0048] The acquisition module is used to acquire the HRRP sequence to be identified.

[0049] The temporal feature extraction module is used to input the HRRP sequence into the temporal feature encoder module of the Trans-FCN model to extract the temporal features of the HRRP sequence.

[0050] The spatial feature extraction module is used to input the HRRP sequence into the spatial feature full convolution module of the Trans-FCN model to extract the spatial features of the HRRP sequence.

[0051] The output module is used to input the temporal features of the HRRP sequence and the spatial features of the HRRP sequence into the decision fusion recognition module of the Trans-FCN model to obtain the radar target recognition result.

[0052] Among them, the HRRP sequence is input into the temporal feature encoder module of the Trans-FCN model to extract the temporal features of the HRRP sequence, including:

[0053] A high-dimensional linear mapping is applied to the HRRP sequence to obtain a high-dimensional sequence; learnable position information is added to the high-dimensional sequence to obtain the input data of the temporal feature encoder; the output of the temporal feature encoder is obtained according to the input data of the temporal feature encoder and the temporal feature encoder; the output of the temporal feature encoder is input into the global average pooling layer of the temporal feature encoder module to obtain the temporal features of the HRRP sequence; wherein the temporal feature encoder includes a grouped linear transformation layer, a self-attention layer, a Dropout regularization layer, a first residual connection layer and a normalization layer, a lightweight feedforward neural network layer, and a second residual connection layer and a normalization layer.

[0054] The spatial feature full convolution module includes a first convolution block, a first squeeze excitation layer, a second convolution block, a second squeeze excitation layer, a third convolution block, and a pooling layer of the spatial feature full convolution module.

[0055] Optionally, the time feature extraction module is further used to:

[0056] S21. Input the temporal feature encoder input data into the grouped linear transformation layer to obtain the self-attention layer input data.

[0057] S22. Input the self-attention layer input data into the self-attention layer to obtain the attention feature.

[0058] S23. Input the attention feature into the Dropout regularization layer, the first residual connection layer and the normalization layer to obtain the first residual connection layer output time feature.

[0059] S24. Input the output time feature of the first residual connection layer into the lightweight feedforward neural network layer to obtain the output time feature of the lightweight feedforward neural network layer.

[0060] S25. Input the temporal features output by the lightweight feedforward neural network layer into the second residual connection layer and the normalization layer to obtain the temporal feature encoder output.

[0061] Optionally, the time feature extraction module is further used to:

[0062] The input data of the temporal feature encoder is grouped to obtain 4 groups of data, each of the 4 groups of data is dimensionally upgraded to obtain 4 groups of high-dimensional features, and the 4 groups of high-dimensional features are merged to obtain a high-dimensional vector.

[0063] The high-dimensional vectors are grouped to obtain 4 groups of data. Each of the 4 groups of data is reduced in dimension to obtain 4 groups of low-dimensional features. The 4 groups of low-dimensional features are concatenated to obtain the input data of the self-attention layer.

[0064] Optionally, the attention feature is as shown in the following formula (1):

[0065] (1)

[0066] In the formula, Represents the attention feature, represents the query vector, Represents the input data of the self-attention layer, represents the key vector, represents a value vector, , , represents the learnable weight matrix.

[0067] Optionally, the spatial feature extraction module is further used to:

[0068] S31. Input the HRRP sequence into the first convolution block to obtain the output spatial features of the first convolution block.

[0069] S32. Input the spatial features output by the first convolution block into the first squeeze excitation layer to obtain the first squeeze excitation layer output.

[0070] S33. Input the output of the first squeeze-excitation layer into the second convolution block to obtain the output spatial features of the second convolution block.

[0071] S34. Input the spatial features output by the second convolution block into the second squeeze excitation layer to obtain the second squeeze excitation layer output.

[0072] S35. Input the output of the second squeeze-excitation layer into the third convolution block to obtain the output spatial features of the third convolution block.

[0073] S36. Input the spatial features output by the third convolution block into the pooling layer of the spatial feature full convolution module to obtain the spatial features of the HRRP sequence.

[0074] Among them, the first convolution block, the second convolution block and the third convolution block respectively include a one-dimensional convolution layer, a batch normalization layer and a ReLU activation function.

[0075] The size of the convolution kernel of the first convolution block is 8, the size of the convolution kernel of the second convolution block is 5, and the size of the convolution kernel of the third convolution block is 3.

[0076] Optionally, the output module is further configured to:

[0077] S41. Input the temporal features of the HRRP sequence and the spatial features of the HRRP sequence into the fully connected layer respectively, and classify them through the softmax function to obtain the temporal branch classification result and the spatial branch classification result.

[0078] S42, establish a learnable time branch decision weight matrix and a space branch decision weight matrix, and obtain a radar target recognition result according to the time branch classification result, the space branch classification result, the time branch decision weight matrix and the space branch decision weight matrix, as shown in the following formula (2):

[0079] (2)

[0080] In the formula, Indicates the radar target recognition result, represents the time branch decision weight matrix, represents the time branch classification result, represents the spatial branch decision weight matrix, Represents the spatial branch classification result.

[0081] Optionally, the process of constructing the loss function of the Trans-FCN model includes:

[0082] Combining the label smoothing strategy with the focal loss function, we get the combined loss function, as shown in the following formula (3):

[0083] (3)

[0084] In the formula, represents the combined loss function, Indicates the number of categories of the true label, represents the smoothing factor, Indicates The one-hot encoded labels of the classes, represents the balance factor, Represents the model for The predicted probability of the class, Represents the adjustment factor.

[0085] The balance factor and adjustment factor in the combined loss function are dynamically adaptively processed to obtain an adaptive focus loss function combined with a label smoothing strategy, as shown in the following formula (4):

[0086] (4)

[0087] In the formula, represents the adaptive focal loss function combined with the label smoothing strategy, represents the total number of samples of all categories, Indicates The number of samples of the class, represents the initial value of the dynamic adjustment factor, represents an adjustable hyperparameter.

[0088] The beneficial effects brought about by the technical solution provided by the embodiment of the present invention include at least:

[0089] In the present invention, the Transformer and the fully convolutional network are combined, and the deep temporal correlation features of the HRRP sequence are extracted by using a lightweight temporal feature encoder, and the deep spatial structure features of the HRRP sequence are extracted by using a spatial feature fully convolutional module. The spatiotemporal features are adaptively fused at the decision level, which improves the robustness and stability of the model while being lightweight, and is conducive to deployment on edge devices. At the same time, in response to the problem of sample category imbalance, an adaptive focal loss function combined with a label smoothing strategy is designed, which can enhance the model's recognition and generalization capabilities for minority samples, and has application value in the recognition scenario of non-cooperative targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0090] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0091] Figure 1 It is a flow chart of a HRRP sequence recognition method based on a lightweight Transformer and a fully convolutional network provided by an embodiment of the present invention;

[0092] Figure 2 is a Trans-FCN network structure diagram provided by an embodiment of the present invention;

[0093] Figure 3 It is a diagram of the construction process of the adaptive focus loss function AFL-LS combined with the label smoothing strategy provided by an embodiment of the present invention;

[0094] Figure 4 It is a block diagram of a HRRP sequence recognition device based on a lightweight Transformer and a fully convolutional network provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0095] The technical solution of the present invention is described below in conjunction with the accompanying drawings.

[0096] In the embodiments of the present invention, words such as "exemplarily" and "for example" are used to indicate examples, illustrations or explanations. Any embodiment or design described as "example" in the present invention should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of the word "example" is intended to present the concept in a specific way. In addition, in the embodiments of the present invention, the meaning expressed by "and / or" can be both, or it can be either of the two.

[0097] In the embodiments of the present invention, "image" and "picture" can sometimes be used interchangeably. It should be noted that when the difference between them is not emphasized, the meanings they intend to express are the same. "of", "corresponding, relevant" and "corresponding" can sometimes be used interchangeably. It should be noted that when the difference between them is not emphasized, the meanings they intend to express are the same.

[0098] In the embodiments of the present invention, sometimes a subscript such as W1 may be written as a non-subscript such as W1. When the difference is not emphasized, the meanings to be expressed are the same.

[0099] In order to make the technical problems, technical solutions and advantages to be solved by the present invention more clear, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0100] The embodiment of the present invention provides a HRRP sequence recognition method based on a lightweight Transformer and a fully convolutional network. The method can be implemented by an HRRP sequence recognition device, which can be a terminal or a server. Figure 1The flowchart of the HRRP sequence recognition method based on the lightweight Transformer and the full convolutional network is shown. The processing flow of the method may include the following steps:

[0101] S1. Obtain a HRRP sequence to be identified.

[0102] In a feasible implementation, the input HRRP sequence is , the dimension of the sequence is ,in, is the length of the HRRP sequence, is the channel dimension of the HRRP sequence, Indicates HRRP data of azimuth angles, Represents matrix transpose.

[0103] S2. Input the HRRP sequence into the temporal feature encoder module of the Trans-FCN model to extract the temporal features of the HRRP sequence.

[0104] Among them, the temporal feature encoder module in S2 inputs the HRRP sequence into the Trans-FCN model to extract the temporal features of the HRRP sequence, including:

[0105] A high-dimensional linear mapping is applied to the HRRP sequence to obtain a high-dimensional sequence; learnable position information is added to the high-dimensional sequence to obtain the input data of the temporal feature encoder; the output of the temporal feature encoder is obtained according to the input data of the temporal feature encoder and the temporal feature encoder; the output of the temporal feature encoder is input into the global average pooling layer of the temporal feature encoder module to obtain the temporal features of the HRRP sequence; wherein the temporal feature encoder includes a grouped linear transformation layer, a self-attention layer, a Dropout regularization layer, a first residual connection layer and a normalization layer, a lightweight feedforward neural network layer, and a second residual connection layer and a normalization layer.

[0106] In one feasible implementation, for the time branch, a high-dimensional linear mapping is first used to map the channel dimension of the sequence from c to Dimension, calculated as:

[0107] (5)

[0108] in, is the sequence after linear mapping, and the channel dimension is ; is a linear mapping matrix; aggregating information into a high-dimensional space is conducive to feature extraction by the self-attention mechanism.

[0109] Furthermore, in order to increase the learnable position information, the high-dimensional sequence containing position information The calculation method is:

[0110] (6)

[0111] in, is the position information matrix, express -dimensional real matrix.

[0112] Furthermore, the temporal feature encoder module is As input, the features are extracted by a Transformer encoder and the temporal features are obtained by dimensionality reduction through a global average pooling layer. .

[0113] Optionally, obtaining a time feature encoder output according to the time feature encoder input data and the time feature encoder includes:

[0114] S21. Input the temporal feature encoder input data into the grouped linear transformation layer to obtain the self-attention layer input data.

[0115] Optionally, in S21, inputting the temporal feature encoder input data into the grouped linear transformation layer to obtain the self-attention layer input data includes:

[0116] The input data of the temporal feature encoder is grouped to obtain 4 groups of data, each of the 4 groups of data is dimensionally upgraded to obtain 4 groups of high-dimensional features, and the 4 groups of high-dimensional features are merged to obtain a high-dimensional vector.

[0117] The high-dimensional vectors are grouped to obtain 4 groups of data. Each of the 4 groups of data is reduced in dimension to obtain 4 groups of low-dimensional features. The 4 groups of low-dimensional features are concatenated to obtain the input data of the self-attention layer.

[0118] In a feasible implementation, the grouped linear transformation layer is divided into two operations: dimension increase and dimension reduction. First, the input data Divided into 4 groups, the data dimension of each group is , mapped to a high-dimensional space by group to obtain high-dimensional features that are easier to distinguish. The process is:

[0119] (7)

[0120] in, is the learnable weight matrix, is the number of channels in the high-dimensional space, is the bias term of the dimension-raising operation, .

[0121] Furthermore, after group-wise dimension upgrading, Merge, and the data dimension is High-dimensional vector , the ReLU activation function is used to increase the nonlinear expression ability of the model, expressed as:

[0122] (8)

[0123] Then, we perform dimensionality reduction by group, and divide them into 4 groups and map them to low-dimensional space. We concatenate the low-dimensional features and use them as the input of the self-attention layer. The process is as follows:

[0124] (9)

[0125] in, is the learnable weight matrix, is the number of channels in the low-dimensional space, is the bias term of the dimensionality reduction operation, . After dimensionality reduction by group Merge, and the data dimension is Vector As input to the self-attention layer.

[0126] S22. Input the self-attention layer input data into the self-attention layer to obtain the attention feature.

[0127] In a feasible implementation, the self-attention layer calculates the correlation of sequence data through the query vector and the key vector to obtain an attention matrix, perceives the global correlation of the sequence, and obtains the attention feature:

[0128] (1)

[0129] In the formula, Represents the attention feature, represents the query vector, Represents the input data of the self-attention layer, represents the key vector, represents a value vector, , , represents the learnable weight matrix.

[0130] S23. Input the attention feature into the Dropout regularization layer, the first residual connection layer and the normalization layer to obtain the first residual connection layer output time feature.

[0131] In a feasible implementation, in order to keep the input and output dimensions consistent and avoid the loss of original features due to the model being too deep, it is necessary to perform dimension increase, residual connection and normalization on the attention feature vector A to obtain the feature :

[0132] (10)

[0133] in, is the parameter matrix for increasing the dimension of the attention feature A. Representation layer normalization function.

[0134] S24. Input the output time feature of the first residual connection layer into the lightweight feedforward neural network layer to obtain the output time feature of the lightweight feedforward neural network layer.

[0135] In one feasible implementation, in a lightweight feedforward neural network, assuming that the dimension of the input vector is , first reduce its dimension to Then, upgrade to Compared with the non-lightweight feedforward neural network, the number of lightweight parameters is only 1 / 2 of the original one while the performance remains unchanged. , the features extracted by the lightweight feedforward neural network are:

[0136] (11)

[0137] in, and is the parameter matrix and bias vector during dimensionality reduction, and are the parameter matrix and bias vector when the dimension is increased, represents a non-linear activation function.

[0138] S25. Input the temporal features output by the lightweight feedforward neural network layer into the second residual connection layer and the normalization layer to obtain the temporal feature encoder output.

[0139] In a feasible implementation, the extracted features are subjected to residual connection and normalization processing, and finally passed through a pooling layer to obtain the temporal features. :

[0140] (12)

[0141] in, represents the global average pooling layer, Representation layer normalization function.

[0142] S3. Input the HRRP sequence into the spatial feature full convolution module of the Trans-FCN model to extract the spatial features of the HRRP sequence.

[0143] The spatial feature full convolution module includes a first convolution block, a first squeeze excitation layer, a second convolution block, a second squeeze excitation layer, a third convolution block, and a pooling layer of the spatial feature full convolution module.

[0144] In a feasible implementation, for the spatial branch, the spatial feature full convolution module is As input, multiple stacked one-dimensional convolutional layers, batch normalization layers, squeeze-excitation layers, and pooling layers are used to extract features to obtain spatial features. .

[0145] Optionally, in S3, the HRRP sequence is input into the spatial feature full convolution module of the Trans-FCN model to extract the spatial features of the HRRP sequence, including:

[0146] S31. Input the HRRP sequence into the first convolution block to obtain the output spatial features of the first convolution block.

[0147] S32. Input the spatial features output by the first convolution block into the first squeeze excitation layer to obtain the first squeeze excitation layer output.

[0148] S33. Input the output of the first squeeze-excitation layer into the second convolution block to obtain the output spatial features of the second convolution block.

[0149] S34. Input the spatial features output by the second convolution block into the second squeeze excitation layer to obtain the second squeeze excitation layer output.

[0150] S35. Input the output of the second squeeze-excitation layer into the third convolution block to obtain the output spatial features of the third convolution block.

[0151] S36. Input the spatial features output by the third convolution block into the pooling layer of the spatial feature full convolution module to obtain the spatial features of the HRRP sequence.

[0152] Among them, the first convolution block, the second convolution block and the third convolution block respectively include a one-dimensional convolution layer, a batch normalization layer and a ReLU activation function.

[0153] The size of the convolution kernel of the first convolution block is 8, the size of the convolution kernel of the second convolution block is 5, and the size of the convolution kernel of the third convolution block is 3.

[0154] In a feasible implementation, the spatial feature full convolution module consists of a one-dimensional convolution layer, a batch normalization layer, a squeeze-excitation layer and a pooling layer, wherein the convolution block consisting of the one-dimensional convolution layer, the batch normalization layer and the ReLU activation function is used to extract the spatial features of the input; the spatial feature full convolution module contains three convolution blocks, and the sizes of the convolution kernels are 8, 5 and 3, respectively, which can capture spatial information of different scales and further improve the accuracy of target recognition; squeeze-excitation layers are added after the first two convolution blocks, and the squeeze-excitation layers perform feature recalibration through global average pooling and fully connected layers to enhance the representation ability of the model. The spatial feature extraction process can be expressed as:

[0155] (13)

[0156] in, represents the squeeze excitation layer, represents a nonlinear activation function, represents the batch normalization layer, represents a one-dimensional convolutional layer; Represents the global average pooling layer, which reduces the dimension of the output of the convolutional layer to obtain the global spatial features .

[0157] S4. Input the temporal features of the HRRP sequence and the spatial features of the HRRP sequence into the decision fusion recognition module of the Trans-FCN model to obtain the radar target recognition result.

[0158] Optionally, in S4, the temporal features of the HRRP sequence and the spatial features of the HRRP sequence are input into a decision fusion recognition module of the Trans-FCN model to obtain a radar target recognition result, including:

[0159] S41. Input the temporal features of the HRRP sequence and the spatial features of the HRRP sequence into the fully connected layer respectively, and classify them through the softmax function to obtain the temporal branch classification result and the spatial branch classification result.

[0160] S42, establishing a learnable time branch decision weight matrix and a space branch decision weight matrix, and obtaining a radar target recognition result according to the time branch classification result, the space branch classification result, the time branch decision weight matrix and the space branch decision weight matrix.

[0161] In a feasible implementation, the decision recognition fusion module is composed of two classifiers, time recognition weights, space recognition weights and a softmax function; the two classifiers use a fully connected layer to respectively classify the time features. and spatial characteristics Perform identification and obtain time identification results And spatial recognition results , establish a learnable decision weight matrix and , ,in is the number of target categories, and the classifier outputs of temporal features and spatial features are adaptively fused. The final recognition result calculation process is:

[0162] (2)

[0163] In the formula, Indicates the radar target recognition result, represents the time branch decision weight matrix, represents the time branch classification result, represents the spatial branch decision weight matrix, represents the spatial branch classification result, , the decision weights can be adaptively integrated at the category level, is the number of classification categories; finally, softmax is used to constrain the output value of the final recognition result.

[0164] The present invention designs an adaptive focus loss function AFL-LS combined with a label smoothing strategy. The specific construction process is as follows: Figure 2 As shown in the figure. The loss function AFL-LS first improves the standard focal loss function and adjusts the balance factor and regulatory factors Perform dynamic adaptive processing so that it can be dynamically adjusted according to the number of categories and prediction probability.

[0165] Specifically, the calculation process of Focal loss (focus loss function) is:

[0166] (14)

[0167] in, represents the predicted probability of the true class of the sample, is a balancing factor used to balance the impact of positive and negative samples. It is the adjustment factor for adjusting the attention of difficult and easy samples.

[0168] Furthermore, label smoothing is a regularization technique used to prevent the model from making overconfident predictions about the training data, thereby improving the generalization ability of the model. The core idea of ​​label smoothing is to convert the hard-coded vector of the true label into a smoother distribution. This method can effectively reduce the overfitting of the model and enhance its robustness to noisy data. Assume that the true label is , the number of categories is , then the one-hot encoding is (Only the The position is 1, and the other positions are 0). After label smoothing It can be expressed as:

[0169] (15)

[0170] in, is the smoothing factor. Usually, label smoothing is combined with the cross entropy loss function (CE), and its calculation expression is:

[0171] (16)

[0172] in, Represents the model for The predicted probability of the class. In the method proposed in this application, the label smoothing strategy is combined with Focal Loss, and the calculation expression is:

[0173] (3)

[0174] The smooth target distribution provided by label smoothing makes the loss function smoother in predicting each category, while FocalLoss makes the model more sensitive to difficult samples by adjusting the loss weight. The combination of the two can improve the model's robustness to noisy data while enhancing the recognition ability of minority class samples.

[0175] For the balance factor in the standard Focal Loss and regulatory factors The value of is usually fixed, and the size of the set value will have an important impact on the recognition effect of difficult samples. In the present invention, the balance factor and regulatory factors Dynamic adaptive processing is performed so that it can be dynamically adjusted according to the number of categories and predicted probability. For each category, the balance factor The calculation formula is:

[0176] (17)

[0177] in, is the total number of samples of all categories, Indicates For the minority sample category, the balance factor A larger value can make the model pay more attention to the minority class. Class adjustment factor , the calculation process is:

[0178] (18)

[0179] in, represents the initial value of the dynamic adjustment factor, is an adjustable hyperparameter, Indicates the prediction for For samples that are more difficult to classify, Small sample size, adjustment factor will be larger, thus giving more attention to such samples. Combining formula (3), the calculation formula of AFL-LS of the present invention can be obtained as follows:

[0180] (4)

[0181] The Trans-FCN (Transformer-Fully Convolutional Networks) proposed in the present invention is a HRRP sequence recognition method based on a lightweight Transformer and a fully convolutional network. The overall structure is a dual-branch structure network, including a temporal feature encoder module, a spatial feature fully convolutional module, and a decision fusion recognition module. The temporal feature encoder module and the spatial feature fully convolutional module are used to extract the temporal and spatial features of the HRRP target sequence respectively; the encoder uses a grouped linear transformation method and a lightweight feedforward neural network for lightweight processing, the convolution extraction module uses a fully convolutional network and a squeeze excitation method to enhance the ability to extract target spatial correlation information, and the decision fusion recognition module performs adaptive fusion recognition on the decision results of temporal and spatial features.

[0182] At the same time, an adaptive focus loss function combined with a label smoothing strategy is designed for use on the class-imbalanced HRRP dataset, which can enhance the model's attention to the minority class while improving the model's generalization ability. The present invention uses a temporal feature encoder module and a spatial feature convolution module to extract the temporal and spatial correlation features of the HRRP target sequence respectively, and at the same time makes lightweight improvements to the Transformer, so that the model has lower computational complexity and model parameters while maintaining a high accuracy rate, which is conducive to the deployment of edge devices and improves the recognition performance on the class-imbalanced HRRP dataset.

[0183] In the embodiment of the present invention, the Transformer and the fully convolutional network are combined, and the deep temporal correlation features of the HRRP sequence are extracted by using a lightweight temporal feature encoder, and the deep spatial structure features of the HRRP sequence are extracted by using a spatial feature fully convolutional module. The spatiotemporal features are adaptively fused at the decision level, which improves the robustness and stability of the model while making it lightweight, and is conducive to deployment on edge devices. At the same time, in response to the problem of sample category imbalance, an adaptive focal loss function combined with a label smoothing strategy is designed, which can enhance the model's recognition and generalization capabilities for minority samples, and has application value in the recognition scenario of non-cooperative targets.

[0184] Figure 4 The block diagram of a HRRP sequence recognition device based on a lightweight Transformer and a full convolutional network is shown according to an exemplary embodiment. The device is used for a HRRP sequence recognition method based on a lightweight Transformer and a full convolutional network. Figure 4 The device includes an acquisition module 310, a time feature extraction module 320, a space feature extraction module 330 and an output module 340. Among them:

[0185] The acquisition module 310 is used to acquire the HRRP sequence to be identified.

[0186] The time feature extraction module 320 is used to input the HRRP sequence into the time feature encoder module of the Trans-FCN model to extract the time features of the HRRP sequence.

[0187] The spatial feature extraction module 330 is used to input the HRRP sequence into the spatial feature full convolution module of the Trans-FCN model to extract the spatial features of the HRRP sequence.

[0188] The output module 340 is used to input the time characteristics of the HRRP sequence and the spatial characteristics of the HRRP sequence into the decision fusion recognition module of the Trans-FCN model to obtain the radar target recognition result.

[0189] Among them, the HRRP sequence is input into the temporal feature encoder module of the Trans-FCN model to extract the temporal features of the HRRP sequence, including:

[0190] A high-dimensional linear mapping is applied to the HRRP sequence to obtain a high-dimensional sequence; learnable position information is added to the high-dimensional sequence to obtain the input data of the temporal feature encoder; the output of the temporal feature encoder is obtained according to the input data of the temporal feature encoder and the temporal feature encoder; the output of the temporal feature encoder is input into the global average pooling layer of the temporal feature encoder module to obtain the temporal features of the HRRP sequence; wherein the temporal feature encoder includes a grouped linear transformation layer, a self-attention layer, a Dropout regularization layer, a first residual connection layer and a normalization layer, a lightweight feedforward neural network layer, and a second residual connection layer and a normalization layer.

[0191] The spatial feature full convolution module includes a first convolution block, a first squeeze excitation layer, a second convolution block, a second squeeze excitation layer, a third convolution block, and a pooling layer of the spatial feature full convolution module.

[0192] In the embodiment of the present invention, the Transformer and the fully convolutional network are combined, and the deep temporal correlation features of the HRRP sequence are extracted by using a lightweight temporal feature encoder, and the deep spatial structure features of the HRRP sequence are extracted by using a spatial feature fully convolutional module. The spatiotemporal features are adaptively fused at the decision level, which improves the robustness and stability of the model while making it lightweight, and is conducive to deployment on edge devices. At the same time, in response to the problem of sample category imbalance, an adaptive focal loss function combined with a label smoothing strategy is designed, which can enhance the model's recognition and generalization capabilities for minority samples, and has application value in the recognition scenario of non-cooperative targets.

[0193] The above embodiments can be implemented in whole or in part by software, hardware (such as circuits), firmware or any other combination. When implemented by software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the process or function described in the embodiment of the present invention is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center by wired (such as infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that contains one or more available media sets. The available medium can be a magnetic medium (for example, a floppy disk, a hard disk, a tape), an optical medium (for example, a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state hard disk.

[0194] It should be understood that the term "and / or" in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. A and B can be singular or plural. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship, but it may also indicate an "and / or" relationship. Please refer to the context for specific understanding.

[0195] In the present invention, "at least one" means one or more, and "plurality" means two or more. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single items or plural items. For example, at least one of a, b, or c can mean: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or multiple.

[0196] It should be understood that in various embodiments of the present invention, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0197] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0198] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described equipment, devices and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0199] In the several embodiments provided by the present invention, it should be understood that the disclosed devices, apparatuses and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another device, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0200] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0201] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0202] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, and other media that can store program codes.

[0203] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A HRRP sequence recognition method based on lightweight Transformer and full convolutional network, characterized in that: The method comprises: S1. Obtain the HRRP sequence to be identified; S2, inputting the HRRP sequence into the temporal feature encoder module of the Trans-FCN model to extract the temporal features of the HRRP sequence; S3, inputting the HRRP sequence into the spatial feature full convolution module of the Trans-FCN model to extract the spatial features of the HRRP sequence; S4, inputting the temporal features of the HRRP sequence and the spatial features of the HRRP sequence into the decision fusion recognition module of the Trans-FCN model to obtain a radar target recognition result; The step of inputting the HRRP sequence into the temporal feature encoder module of the Trans-FCN model in S2 to extract the temporal features of the HRRP sequence includes: Apply high-dimensional linear mapping to the HRRP sequence to obtain a high-dimensional sequence; add learnable position information to the high-dimensional sequence to obtain time feature encoder input data; obtain a time feature encoder output according to the time feature encoder input data and the time feature encoder; input the time feature encoder output to the global average pooling layer of the time feature encoder module to obtain the time feature of the HRRP sequence; wherein the time feature encoder includes a grouped linear transformation layer, a self-attention layer, a Dropout regularization layer, a first residual connection layer and a normalization layer, a lightweight feedforward neural network layer, and a second residual connection layer and a normalization layer; The spatial feature full convolution module includes a first convolution block, a first squeeze excitation layer, a second convolution block, a second squeeze excitation layer, a third convolution block and a pooling layer of the spatial feature full convolution module.

2. The HRRP sequence recognition method based on lightweight Transformer and full convolutional network according to claim 1 is characterized in that: The step of obtaining the time feature encoder output according to the time feature encoder input data and the time feature encoder comprises: S21, inputting the temporal feature encoder input data into the grouped linear transformation layer to obtain self-attention layer input data; S22, inputting the self-attention layer input data into the self-attention layer to obtain attention features; S23, inputting the attention feature into the Dropout regularization layer, the first residual connection layer and the normalization layer to obtain the first residual connection layer output time feature; S24, inputting the first residual connection layer output time feature into the lightweight feedforward neural network layer to obtain the lightweight feedforward neural network layer output time feature; S25. Input the temporal features output by the lightweight feedforward neural network layer into the second residual connection layer and the normalization layer to obtain a temporal feature encoder output.

3. The HRRP sequence recognition method based on lightweight Transformer and full convolutional network according to claim 2 is characterized in that: The step of inputting the temporal feature encoder input data into the grouped linear transformation layer to obtain self-attention layer input data in S21 includes: The input data of the temporal feature encoder is grouped to obtain 4 groups of data, each of the 4 groups of data is dimensionally upgraded to obtain 4 groups of high-dimensional features, and the 4 groups of high-dimensional features are combined to obtain a high-dimensional vector; The high-dimensional vectors are grouped to obtain 4 groups of data, each of the 4 groups of data is reduced in dimension to obtain 4 groups of low-dimensional features, and the 4 groups of low-dimensional features are concatenated to obtain self-attention layer input data.

4. The HRRP sequence recognition method based on lightweight Transformer and full convolutional network according to claim 2 is characterized in that: The attention feature in S22 is shown in the following formula (1): （1） In the formula, Represents the attention feature, represents the query vector, Represents the input data of the self-attention layer, represents the key vector, represents a value vector, , , represents the learnable weight matrix.

5. The HRRP sequence recognition method based on lightweight Transformer and full convolutional network according to claim 1 is characterized in that: The step of inputting the HRRP sequence into the spatial feature full convolution module of the Trans-FCN model in S3 to extract the spatial features of the HRRP sequence includes: S31, inputting the HRRP sequence into the first convolution block to obtain the output spatial features of the first convolution block; S32, inputting the spatial features output by the first convolution block into the first squeeze excitation layer to obtain a first squeeze excitation layer output; S33, inputting the first squeeze excitation layer output into the second convolution block to obtain the second convolution block output spatial features; S34, inputting the second convolution block output spatial features into the second squeeze excitation layer to obtain a second squeeze excitation layer output; S35, inputting the output of the second squeeze-excitation layer into the third convolution block to obtain the output spatial features of the third convolution block; S36, inputting the spatial features output by the third convolution block into the pooling layer of the spatial feature full convolution module to obtain the spatial features of the HRRP sequence; The first convolution block, the second convolution block and the third convolution block respectively include a one-dimensional convolution layer, a batch normalization layer and a ReLU activation function; The size of the convolution kernel of the first convolution block is 8, the size of the convolution kernel of the second convolution block is 5, and the size of the convolution kernel of the third convolution block is 3.

6. The HRRP sequence recognition method based on lightweight Transformer and full convolutional network according to claim 1 is characterized in that: The step of inputting the temporal features of the HRRP sequence and the spatial features of the HRRP sequence into the decision fusion recognition module of the Trans-FCN model in S4 to obtain the radar target recognition result includes: S41, respectively inputting the temporal features of the HRRP sequence and the spatial features of the HRRP sequence into the fully connected layer, and performing classification through the softmax function to obtain a temporal branch classification result and a spatial branch classification result; S42, establish a learnable time branch decision weight matrix and a space branch decision weight matrix, and obtain a radar target recognition result according to the time branch classification result, the space branch classification result, the time branch decision weight matrix and the space branch decision weight matrix, as shown in the following formula (2): （2） In the formula, Indicates the radar target recognition result, represents the time branch decision weight matrix, represents the time branch classification result, represents the spatial branch decision weight matrix, Represents the spatial branch classification result.

7. The HRRP sequence recognition method based on lightweight Transformer and full convolutional network according to claim 1 is characterized in that: The process of constructing the loss function of the Trans-FCN model includes: Combining the label smoothing strategy with the focal loss function, we get the combined loss function, as shown in the following formula (3): （3） In the formula, represents the combined loss function, Indicates the number of categories of the true label, represents the smoothing factor, Indicates The one-hot encoded labels of the classes, represents the balance factor, Represents the model for The predicted probability of the class, represents the regulating factor; The balance factor and the adjustment factor in the combined loss function are dynamically adaptively processed to obtain an adaptive focus loss function combined with a label smoothing strategy, as shown in the following formula (4): （4） In the formula, represents the adaptive focal loss function combined with the label smoothing strategy, represents the total number of samples of all categories, Indicates The number of samples of the class, represents the initial value of the dynamic adjustment factor, represents an adjustable hyperparameter.

8. A HRRP sequence recognition device based on a lightweight Transformer and a full convolutional network, wherein the HRRP sequence recognition device based on a lightweight Transformer and a full convolutional network is used to implement the HRRP sequence recognition method based on a lightweight Transformer and a full convolutional network as described in any one of claims 1 to 7, characterized in that: The device comprises: An acquisition module, used for acquiring a HRRP sequence to be identified; A time feature extraction module, used for inputting the HRRP sequence into the time feature encoder module of the Trans-FCN model to extract the time feature of the HRRP sequence; A spatial feature extraction module, used for inputting the HRRP sequence into a spatial feature full convolution module of a Trans-FCN model to extract spatial features of the HRRP sequence; An output module, used for inputting the temporal features of the HRRP sequence and the spatial features of the HRRP sequence into a decision fusion recognition module of a Trans-FCN model to obtain a radar target recognition result; The step of inputting the HRRP sequence into a temporal feature encoder module of a Trans-FCN model to extract the temporal features of the HRRP sequence includes: Apply high-dimensional linear mapping to the HRRP sequence to obtain a high-dimensional sequence; add learnable position information to the high-dimensional sequence to obtain time feature encoder input data; obtain a time feature encoder output according to the time feature encoder input data and the time feature encoder; input the time feature encoder output to the global average pooling layer of the time feature encoder module to obtain the time feature of the HRRP sequence; wherein the time feature encoder includes a grouped linear transformation layer, a self-attention layer, a Dropout regularization layer, a first residual connection layer and a normalization layer, a lightweight feedforward neural network layer, and a second residual connection layer and a normalization layer; The spatial feature full convolution module includes a first convolution block, a first squeeze excitation layer, a second convolution block, a second squeeze excitation layer, a third convolution block and a pooling layer of the spatial feature full convolution module.

9. The HRRP sequence recognition device based on lightweight Transformer and full convolutional network according to claim 8, characterized in that: The step of obtaining the time feature encoder output according to the time feature encoder input data and the time feature encoder comprises: S21, inputting the temporal feature encoder input data into the grouped linear transformation layer to obtain self-attention layer input data; S22, inputting the self-attention layer input data into the self-attention layer to obtain attention features; S23, inputting the attention feature into the Dropout regularization layer, the first residual connection layer and the normalization layer to obtain the first residual connection layer output time feature; S24, inputting the first residual connection layer output time feature into the lightweight feedforward neural network layer to obtain the lightweight feedforward neural network layer output time feature; S25. Input the temporal features output by the lightweight feedforward neural network layer into the second residual connection layer and the normalization layer to obtain a temporal feature encoder output.

10. The HRRP sequence recognition device based on lightweight Transformer and full convolutional network according to claim 9, characterized in that: The step of inputting the temporal feature encoder input data into the grouped linear transformation layer to obtain self-attention layer input data in S21 includes: The input data of the temporal feature encoder is grouped to obtain 4 groups of data, each of the 4 groups of data is dimensionally upgraded to obtain 4 groups of high-dimensional features, and the 4 groups of high-dimensional features are combined to obtain a high-dimensional vector; The high-dimensional vectors are grouped to obtain 4 groups of data, each of the 4 groups of data is reduced in dimension to obtain 4 groups of low-dimensional features, and the 4 groups of low-dimensional features are concatenated to obtain self-attention layer input data.

Images