HRRP enhanced preprocessing method based on attention mechanism under interference conditions

Abstract

The present invention discloses a HRRP enhancement preprocessing method based on an attention mechanism under interference conditions, including: generating a first data set; setting the data of several positions in the spectrum signal to 0 to obtain a second data set; obtaining an auxiliary data set based on the second data set; preprocessing the second data set to obtain a third data set; based on a Transformer network model, obtaining interpolation values ​​according to the third data set, a mask matrix and a position matrix; using a loss function to calculate the loss value between the interpolation value and the true value; based on the loss value, using a back propagation algorithm to iteratively update the parameters of the Transformer network model until the loss value converges to obtain a trained Transformer network model; based on the trained Transformer network model, obtaining a reconstruction result according to the target spectrum signal. The reconstruction method provided by the present invention is mainly used to solve the problem that the recurrent neural network in the prior art does not fully utilize the high-resolution range image spectrum signal and the information of the preceding data is missing.

Description

Technical Field

[0001] The present invention belongs to the field of radar technology, and in particular relates to a HRRP enhancement preprocessing method based on an attention mechanism under interference conditions. Background Art

[0002] High-resolution range image is the vector sum of the sub-echoes of the target scattering points obtained by broadband radar projected on the radar ray. It provides the distribution of the target scattering points along the distance direction, which is an important structural feature of the target and is very valuable for target recognition. However, when there is narrowband interference in the environment, the imaging and recognition performance of broadband radar will be seriously degraded or even fail. Considering that narrowband interference has the characteristics of strong energy and easy detection, the imaging and recognition tasks can be completed after eliminating the interference component in the frequency domain and interpolating and reconstructing the spectrum. The key lies in how to interpolate the missing values ​​of the spectrum.

[0003] Currently, a commonly used method is to use recurrent neural networks to interpolate missing values.

[0004] However, recurrent neural networks cannot be calculated in parallel and have a long running time. In addition, the previous data information in the recurrent neural network will decay as the cycle proceeds, so its interpolation performance will still be subject to certain limitations. Summary of the invention

[0005] In order to solve the above problems existing in the prior art, the present invention provides a HRRP enhancement preprocessing method based on the attention mechanism under interference conditions. The technical problem to be solved by the present invention is achieved by the following technical solutions:

[0006] A HRRP enhancement preprocessing method based on an attention mechanism under interference conditions, the HRRP enhancement preprocessing method comprising:

[0007] Step 1: Generate a first data set, wherein the first data set includes one-dimensional high-resolution range image spectrum signals of M categories of targets, and the dimension of each spectrum signal is K×2;

[0008] Step 2, setting the data at several positions in the spectrum signal to 0 to obtain a second data set;

[0009] Step 3: obtaining an auxiliary data set according to the second data set, wherein the auxiliary data set includes a mask matrix and a position matrix of the spectrum signal;

[0010] Step 4: preprocess the second data set to obtain a third data set;

[0011] Step 5: Based on the Transformer network model, interpolation values ​​of missing spectral elements are obtained according to the third data set, the mask matrix and the position matrix;

[0012] Step 6: Calculate the loss value between the interpolated value and the true value using the loss function;

[0013] Step 7: Based on the loss value, the parameters of the Transformer network model are iteratively updated using a back propagation algorithm until the loss value converges to obtain a trained Transformer network model;

[0014] Step 8: Based on the trained Transformer network model, interpolation values ​​are obtained according to the target spectrum signal, and the interpolation values ​​are used to fill the missing values ​​of the corresponding spectrum signal to obtain the reconstructed signal.

[0015] In one embodiment of the present invention, positions where the elements in the mask matrix are 0 indicate missing elements, positions where the elements in the mask matrix are 1 indicate non-missing elements, and the value of each element in the position matrix indicates the serial number of the corresponding spectrum signal element.

[0016] In one embodiment of the present invention, step 4 comprises:

[0017] The amplitude of each spectrum signal in the second data set is normalized by modulo-2 norm to obtain the third data set.

[0018] In one embodiment of the present invention, the Transformer network model includes an Embedding module, a feature extraction module, a context feature extraction module and an interpolation module.

[0019] In one embodiment of the present invention, the Embedding module includes a position encoding layer and a feature fusion layer;

[0020] The feature extraction module includes an input layer and a first fully connected layer;

[0021] The context feature extraction module includes several encoding layers, the encoding layer includes a self-attention layer and a feedforward neural network, the self-attention layer includes a multi-head self-attention layer and a second fully connected layer, and the feedforward neural network includes a third fully connected layer and a fourth fully connected layer;

[0022] The interpolation module includes a fifth fully connected layer and an output layer.

[0023] In one embodiment of the present invention, step 5 comprises:

[0024] Step 5.1, inputting the third data set, the mask matrix, and the position matrix into the Embedding module, wherein the position matrix passes through a position encoding layer to obtain a position encoding matrix;

[0025] Step 5.2, concatenating the third data set, the position encoding matrix, and the mask matrix through the feature fusion layer to obtain a feature fusion matrix;

[0026] Step 5.3, inputting the feature fusion matrix into the feature extraction module to obtain a first spectrum feature;

[0027] Step 5.4: input the first spectrum feature into the context feature extraction module, and pass through all coding layers in sequence to obtain spectrum context features;

[0028] Step 5.5: The spectral context features corresponding to the elements that are 0 in the mask matrix are used by the interpolation module to obtain interpolation values ​​of the missing spectral elements.

[0029] In one embodiment of the present invention, the encoding method of the position encoding layer is:

[0030]

[0031]

[0032] Where PE represents the position encoding matrix, pos represents the sequence number of the spectrum element, i represents the column number of the spectrum element in the position encoding matrix, and d model Represents the dimensions of the row vectors in the position encoding matrix.

[0033] In one embodiment of the present invention, step 6 comprises:

[0034] The loss value between the interpolated value and the true value is calculated using the mean absolute error loss function.

[0035] In one embodiment of the present invention, the mean absolute error loss function is:

[0036]

[0037] Among them, pred represents the imputed value, real represents the true value, and N represents the number of missing values.

[0038] In one embodiment of the present invention, step 8 comprises:

[0039] Step 8.1, obtaining a first target spectrum signal;

[0040] Step 8.2: obtaining a mask matrix and a position matrix of the first target spectrum signal according to the first target spectrum signal;

[0041] Step 8.3, performing modulo-2 norm normalization processing on the amplitude of the first target spectrum signal to obtain a second target spectrum signal;

[0042] Step 8.4: Pass the second target spectrum signal through the Embedding module, feature extraction module, context feature extraction module and interpolation module of the trained Transformer network model in sequence to obtain interpolation values ​​of missing values, and use the interpolation values ​​to fill the missing values ​​of the corresponding spectrum signal to obtain a reconstructed signal.

[0043] Beneficial effects of the present invention:

[0044] First, since the present invention constructs a Transformer network model for interpolating missing values, the Transformer network model encodes the position number of each element in the spectral signal and splices it with the mask matrix and the spectral echo, so that the model can use the position information of each element in the spectral signal to assist the network in interpolating the missing values, making the interpolation result more accurate.

[0045] Second, the present invention uses the Transformer network model to extract the contextual features of each element. Compared with the recurrent neural network, the Transformer network model will pay attention to the elements at all positions when processing the current position element to help the model extract the features of the current position element and the dependency information of the current position data on other position data. It overcomes the problem that the information of the preceding elements in the traditional recurrent neural network will decay as the cycle proceeds and cannot be executed in parallel. Compared with the traditional recurrent neural network, better interpolation performance can be achieved.

[0046] Third, the interpolation method provided by the present invention is mainly used to solve the problem in the prior art that the recurrent neural network does not fully utilize the high-resolution range image spectrum signal and the information of the previous data is missing.

[0047] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a flowchart of a HRRP enhancement preprocessing method based on an attention mechanism under interference conditions provided by an embodiment of the present invention;

[0049] Figure 2 It is a schematic diagram of the structure of a Transformer model for interpolating missing values ​​provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0050] The present invention is further described in detail below with reference to specific embodiments, but the embodiments of the present invention are not limited thereto.

[0051] Embodiment 1

[0052] When narrowband interference exists in the environment, the imaging and recognition performance of broadband radar will be seriously degraded or even fail. Considering that narrowband interference has the characteristics of strong energy and easy detection, the imaging and recognition tasks can be completed after removing the interference component in the frequency domain and interpolating and reconstructing the spectrum. Therefore, how to interpolate and reconstruct the missing values ​​of the spectrum to make it close to the true value is an urgent problem to be solved. To solve this problem, please refer to Figure 1 , Figure 1 1 is a flow chart of a HRRP enhancement preprocessing method based on an attention mechanism under interference conditions provided by an embodiment of the present invention. The present invention proposes a HRRP enhancement preprocessing method based on an attention mechanism under interference conditions, and the HRRP enhancement preprocessing method includes:

[0053] Step 1: Generate a first data set, which includes one-dimensional high-resolution range image spectrum signals of M categories of targets, and the dimension of each spectrum signal is K×2.

[0054] Specifically, one-dimensional high-resolution range image spectrum signals of M categories of targets are extracted to form a training data set (i.e., the first data set), and each category contains at least 2000 one-dimensional high-resolution range image spectrum signals. Since the spectrum signal is complex data, the dimension of each spectrum signal is, for example, 256×2, where M≥1.

[0055] Step 2: Set the data at several positions in the spectrum signal to 0 to obtain a second data set.

[0056] Specifically, some spectrum signals are arbitrarily selected from all spectrum signals, and some position data in the selected spectrum signals are set to 0, which indicates that the spectrum at this position is disturbed and missing. The unchanged spectrum signals and the changed spectrum signals constitute the second data set. This embodiment does not limit the positions and specific numbers set to 0 in the spectrum signals.

[0057] Step 3: Obtain an auxiliary data set according to the second data set, where the auxiliary data set includes a mask matrix and a position matrix of the spectrum signal.

[0058] Specifically, a K×1 mask matrix is ​​generated, in which an element of 0 indicates that the spectrum at the corresponding position is missing due to interference, and an element of 1 indicates that the spectrum element at the corresponding position is not missing, for example, the dimension of the mask matrix is ​​256×1; a K×1 position matrix is ​​generated, in which the value of each element in the position matrix indicates the sequence number of the corresponding spectrum signal element, for example, the first element value is 1, the second element is 2, and so on, for example, the dimension of the position matrix is ​​256×1; the mask matrix and position matrix of all spectrum signals are used as auxiliary data sets.

[0059] Step 4: preprocess the second data set to obtain a third data set.

[0060] Specifically, a modulo-2 norm normalization process is performed on the amplitude of each spectrum signal in the second data set to obtain a third data set, that is, all spectrum signals after the normalization process constitute the third data set.

[0061] Step 5: Based on the Transformer network model, according to the third data set, the mask matrix and the position matrix, the interpolation values ​​of the missing spectral elements are obtained, and finally the reconstructed HRRP frequency domain signal is obtained.

[0062] In this example, see Figure 2 ,The Transformer network model includes an Embedding module, a feature extraction module, a context feature extraction module and an interpolation module.

[0063] Among them, the Embedding module includes a position encoding layer and a feature fusion layer, that is, building an Embedding module including a position encoding layer and a feature fusion layer.

[0064] The feature extraction module includes an input layer and a first fully connected layer, that is, a feature extraction module consisting of an input layer and a first fully connected layer is constructed.

[0065] The context feature extraction module includes several encoding layers, the encoding layers include a self-attention layer and a feedforward neural network, the self-attention layer includes a multi-head self-attention layer and a second fully connected layer, and the feedforward neural network includes a third fully connected layer and a fourth fully connected layer; for example, there are 4 encoding layers in total, namely, the first encoding layer, the second encoding layer, the third encoding layer and the fourth encoding layer.

[0066] The parameters of each layer of the context feature extraction module are set as follows: the dimension size of the query vector (Q), key-value vector (K), and query vector (V) in the multi-head self-attention layer is set to 64, the number of attentions is set to 12, and the weight dimension size of the second fully connected layer in the self-attention layer is set to 768×512; the weight dimension sizes of the third fully connected layer and the fourth fully connected layer in the feedforward neural network are set to 512×1024 and 1024×512, respectively.

[0067] The interpolation module includes a fifth fully connected layer and an output layer, that is, an interpolation module consisting of the fifth fully connected layer and the output layer is constructed;

[0068] The Embedding module, feature extraction module, context feature extraction module and interpolation module are combined into a Transformer network model for interpolating missing values.

[0069] See also Figure 2, the structure of the Transformer network model for interpolating missing values ​​constructed in this embodiment is further described in detail.

[0070] exist Figure 2 The high-resolution range image spectrum signal, mask matrix, and position matrix with missing values ​​are input from the leftmost Embedding module, and the position encoding layer and feature fusion layer constitute the Embedding module. The module composed of the first fully connected layer of 11×512 dimensions is the feature extraction module. The context feature extraction module consists of the encoding layer structure in 4 Transformers, in which the dimension size of the query vector (Q), key value vector (K), and query vector (V) in the self-attention layer is set to 64, the number of attentions is set to 12, and the weight dimension size of the second fully connected layer in the self-attention layer is set to 768×512; the weight dimension size of the third fully connected layer and the fourth fully connected layer in the feedforward neural network is set to 512×1024 and 1024×512 respectively. The interpolation module consists of the fifth fully connected layer of 512×1 dimensions and the output layer, and the output of the interpolation module is the interpolation value of the missing value.

[0071] In a specific embodiment, step 5 may specifically include:

[0072] Step 5.1, input the third data set, the mask matrix, and the position matrix into the Embedding module, wherein the position matrix passes through the position encoding layer to obtain the position encoding matrix;

[0073] Specifically, the encoding method of the position encoding layer is:

[0074]

[0075]

[0076] Where PE represents the position encoding matrix, pos represents the sequence number of the spectrum element, i represents the column number of the spectrum element in the position encoding matrix, and d model Represents the dimensions of the row vectors in the position encoding matrix.

[0077] Step 5.2, concatenate the third data set, the position encoding matrix, and the mask matrix through a feature fusion layer to obtain a feature fusion matrix, such as a 256×11 feature fusion matrix;

[0078] Step 5.3, input the feature fusion matrix into the feature extraction module to obtain the first spectrum feature, such as obtaining the first spectrum feature of 256×512;

[0079] Step 5.4, input the first spectrum feature into the context feature extraction module, and pass through all coding layers in sequence to obtain spectrum context features, for example, pass through the first coding layer, the second coding layer, the third coding layer, and the fourth coding layer in sequence to obtain spectrum context features, and the dimension of the spectrum context features is 256×512;

[0080] Step 5.5: The spectral context features corresponding to the elements that are 0 in the mask matrix are interpolated through the interpolation module to obtain the interpolation values ​​of the missing spectral elements.

[0081] Step 6: Use the loss function to calculate the loss value between the interpolated value and the true value.

[0082] Specifically, the mean absolute error (MAE) loss function is used to calculate the loss value between the interpolated value and the true value. The mean absolute error loss function is:

[0083]

[0084] Among them, pred represents the imputed value, real represents the true value, and N represents the number of missing values.

[0085] Step 7: Based on the loss value, use the back propagation algorithm to iteratively update the parameters of the Transformer network model until the loss value converges to obtain a trained Transformer network model.

[0086] Specifically, the Embedding module, feature extraction module, context feature extraction module, and interpolation module are combined to form a Transformer network model for interpolating missing values. All modules participate in the forward propagation process and the error back propagation process at the same time, and all weight parameters of the Transformer network model for interpolating missing values ​​are iteratively updated.

[0087] Step 8: Based on the trained Transformer network model, interpolation values ​​are obtained according to the target spectrum signal, and the interpolation values ​​are used to fill the missing values ​​of the corresponding spectrum signal to obtain the reconstructed signal.

[0088] Step 8.1: Obtain a first target spectrum signal.

[0089] Specifically, elements in the one-dimensional high-resolution range image spectrum signal (ie, the first target spectrum signal) with interference are set to 0.

[0090] Step 8.2: Obtain a mask matrix and a position matrix of the first target spectrum signal according to the first target spectrum signal.

[0091] Specifically, the mask matrix and position matrix of the first target spectrum signal are obtained using the method in step 3.

[0092] Step 8.3: Perform modulo-2 norm normalization processing on the amplitude of the first target spectrum signal to obtain a second target spectrum signal.

[0093] Step 8.4: Pass the second target spectrum signal through the Embedding module, feature extraction module, context feature extraction module and interpolation module of the trained Transformer network model in sequence to obtain interpolation values ​​of missing values, and use the interpolation values ​​to fill the missing values ​​of the corresponding spectrum signal to obtain the reconstructed signal.

[0094] In order to effectively utilize the target spectrum signal, the present invention constructs a Transformer network model for interpolating missing values ​​including an Embedding module, a feature extraction module, a context feature extraction module, and an interpolation module. The spectrum signal, the position sequence number, and the mask matrix are fused through the Embedding module, and then the output is sequentially passed through the feature extraction module, the context feature extraction module, and the interpolation module for preliminary feature extraction, context feature extraction, and interpolation of missing values. By constructing a Transformer network model for interpolating missing data, its operation is simple and the interpolation effect is good.

[0095] The present invention can solve the problem of partial spectrum loss of high-resolution range image spectrum due to interference, and interpolate the missing spectrum of different targets moving in the air and on the ground.

[0096] Since the present invention constructs a Transformer network model for interpolating missing values, the Transformer network model includes an Embedding module, a feature extraction module, a context feature extraction module, and an interpolation module. In the Embedding module, the model encodes the position number of each element in the spectrum signal, and splices it with the mask matrix and the spectrum echo, so that the model can use the position information of each element in the spectrum signal to assist the network in interpolating the missing values, making the interpolation result more accurate.

[0097] Since the present invention uses the Transformer network model to extract the context features of each element in the context feature extraction module, compared to the recurrent neural network, the Transformer network model will pay attention to the elements at all positions when processing the current position element to help the model extract the features of the current position element and the dependency information of the current position data on other position data, thereby overcoming the problems in the traditional recurrent neural network that the information of the preceding elements will decay as the cycle proceeds and cannot be executed in parallel, and better interpolation performance can be achieved compared to the traditional recurrent neural network.

[0098] 1. Simulation experiment conditions:

[0099] The hardware platform of the simulation experiment of the present invention is: the processor is Intel (R) Core (TM) i7-7700 CPU @ 3.60 GHZ 3.60 GHZ, the main frequency is 2.00 GHz, and the memory is 16 GB.

[0100] The software platform for the simulation experiment of the present invention is: Windows 10 operating system and Python 3.6.

[0101] The data used in the experiment are measured data of high-resolution range images of three types of aircraft, namely, Citation (715), An-26 (507), and Yak-42 (922). The three types of high-resolution range image spectrum signals obtained are the high-resolution range image spectrum signals of the Citation (715), the high-resolution range image spectrum signals of the An-26 (507), and the high-resolution range image spectrum signals of the Yak-42 (922). The three types of high-resolution range imaging data are divided into a training sample set and a test sample set, and then corresponding category labels are added to all the high-resolution range imaging data in the training sample set and the test sample set respectively; the training sample set contains 9000 training samples, and the test sample set contains 3000 test samples, among which the training sample set contains 3000 high-resolution range image spectrum signals of the first type, 3000 high-resolution range image spectrum signals of the second type, and 3000 high-resolution range image spectrum signals of the third type. The test sample set contains 1000 high-resolution range image spectrum signals of the first category, 1000 high-resolution range image spectrum signals of the second category, and 1000 high-resolution range image spectrum signals of the third category.

[0102] Before missing value interpolation, the elements at random positions in the original spectrum signal are set to zero, then normalized, and finally the missing value interpolation is performed using the Transformer model to obtain the reconstructed HRRP frequency domain data.

[0103] 2. Experimental content and results analysis

[0104] The simulation experiment of the present invention adopts the present invention and a bidirectional recurrent neural network (BRNN) interpolation method to respectively perform simulation experiments on one-dimensional high-resolution range image spectrum signals of three types of aircraft targets.

[0105] The interpolation results of the two methods were evaluated using relative reconstruction error.

[0106] All samples in the training data set and the auxiliary data set are input into the Transformer network model for interpolating missing values, and one round of training of the Transformer network model for interpolating missing values ​​is completed. The iteration round is recorded as one, and each round of iteration ends, with a total of 200 iterations. After the iteration is completed, all samples in the test data set and the test auxiliary data set are input into the iteratively trained network to obtain the interpolated values ​​of the missing values ​​in the test data set. The interpolated values ​​are used to fill the missing samples to obtain the reconstructed samples. The reconstruction error MRE between the reconstructed samples and the true samples is calculated as follows:

[0107]

[0108] in represents the frequency domain signal of the reconstructed samples, represents the frequency domain signal of the real sample corresponding to the reconstructed sample, and I represents the number of samples to be reconstructed.

[0109] Table 1 shows the relative reconstruction error results of various models when the missing rate of the data set is 10%, 30%, and 50%.

[0110] Table 1

[0111]

[0112]

[0113] As shown in Table 1, the method proposed in the present invention has a significant decrease in the relative reconstruction error of the reconstructed samples compared to the bidirectional recurrent neural network interpolation method. When the missing rate is 10%, 30%, and 50%, the relative reconstruction error of the method proposed in the present invention is lower than that of the bidirectional recurrent neural network. When the missing rate is 50%, the relative reconstruction error of the method proposed in the present invention is 0.1397, while that of the bidirectional recurrent neural network is 0.3247, and the relative reconstruction error decreases by 0.185. At the same time, under the conditions of 10%, 30%, and 50% missing rates, the performance of the method proposed in the present invention is relatively stable, while the performance of the bidirectional recurrent neural network interpolation method decreases significantly. By comparing the relative reconstruction errors under the conditions of 10%, 30%, and 50% missing rates, it can be seen that the method proposed in the present invention has better reconstruction performance and stability than the bidirectional recurrent neural network interpolation method.

[0114] The HRRP enhancement preprocessing method based on the attention mechanism under interference conditions proposed in the present invention has the characteristics of good stability and high interpolation performance. Therefore, the present invention can address the problem of missing high-resolution range spectrum under interference conditions and realize the interpolation and reconstruction of missing spectrum signals, which has important practical significance.

[0115] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine different embodiments or examples described in this specification.

[0116] Although the present application is described herein in conjunction with various embodiments, in the process of implementing the claimed application, those skilled in the art may understand and implement other variations of the disclosed embodiments by viewing the drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other components or steps, and "one" or "an" does not exclude multiple situations. A single processor or other unit may implement several functions listed in a claim. Certain measures are recorded in different dependent claims, but this does not mean that these measures cannot be combined to produce good results.

[0117] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

Claims

1. A HRRP enhancement preprocessing method based on attention mechanism under interference conditions, characterized in that: The HRRP enhanced preprocessing method comprises: Step 1: Generate a first data set, wherein the first data set includes one-dimensional high-resolution range image spectrum signals of M categories of targets, and the dimension of each spectrum signal is K×2; Step 2, setting the data at several positions in the spectrum signal to 0 to obtain a second data set; Step 3: obtaining an auxiliary data set according to the second data set, wherein the auxiliary data set includes a mask matrix and a position matrix of the spectrum signal; Step 4: preprocess the second data set to obtain a third data set; Step 5: Based on the Transformer network model, interpolation values ​​of missing spectral elements are obtained according to the third data set, the mask matrix and the position matrix; Step 6: Calculate the loss value between the interpolated value and the true value using the loss function; Step 7: Based on the loss value, the parameters of the Transformer network model are iteratively updated using a back propagation algorithm until the loss value converges to obtain a trained Transformer network model; Step 8: Based on the trained Transformer network model, interpolation values ​​are obtained according to the target spectrum signal, and the interpolation values ​​are used to fill the missing values ​​of the corresponding spectrum signal to obtain the reconstructed signal.

2. The HRRP enhancement preprocessing method according to claim 1, characterized in that: Positions where the elements in the mask matrix are 0 indicate that the elements are missing, positions where the elements in the mask matrix are 1 indicate that the elements are not missing, and the value of each element in the position matrix indicates the serial number of the corresponding spectrum signal element.

3. The HRRP enhancement preprocessing method according to claim 1, characterized in that: The step 4 comprises: The amplitude of each spectrum signal in the second data set is normalized by modulo-2 norm to obtain the third data set.

4. The HRRP enhancement preprocessing method according to claim 1, characterized in that: The Transformer network model includes an Embedding module, a feature extraction module, a context feature extraction module and an interpolation module.

5. The HRRP enhancement preprocessing method according to claim 4, characterized in that: The Embedding module includes a position encoding layer and a feature fusion layer; The feature extraction module includes an input layer and a first fully connected layer; The context feature extraction module includes several encoding layers, the encoding layer includes a self-attention layer and a feedforward neural network, the self-attention layer includes a multi-head self-attention layer and a second fully connected layer, and the feedforward neural network includes a third fully connected layer and a fourth fully connected layer; The interpolation module includes a fifth fully connected layer and an output layer.

6. The HRRP enhancement preprocessing method according to claim 5, characterized in that: The step 5 comprises: Step 5.1, inputting the third data set, the mask matrix, and the position matrix into the Embedding module, wherein the position matrix passes through a position encoding layer to obtain a position encoding matrix; Step 5.2, concatenating the third data set, the position encoding matrix, and the mask matrix through the feature fusion layer to obtain a feature fusion matrix; Step 5.3, inputting the feature fusion matrix into the feature extraction module to obtain a first spectrum feature; Step 5.4: input the first spectrum feature into the context feature extraction module, and pass through all coding layers in sequence to obtain spectrum context features; Step 5.5: The spectral context features corresponding to the elements that are 0 in the mask matrix are used by the interpolation module to obtain interpolation values ​​of the missing spectral elements.

7. The HRRP enhancement preprocessing method according to claim 6, characterized in that: The encoding method of the position encoding layer is: Where PE represents the position encoding matrix, pos represents the sequence number of the spectrum element, i represents the column number of the spectrum element in the position encoding matrix, and d model Represents the dimensions of the row vectors in the position encoding matrix.

8. The HRRP enhancement preprocessing method according to claim 1, characterized in that: The step 6 comprises: The loss value between the interpolated value and the true value is calculated using the mean absolute error loss function.

9. The HRRP enhancement preprocessing method according to claim 8, characterized in that: The mean absolute error loss function is: Among them, pred represents the imputed value, real represents the true value, and N represents the number of missing values.

10. The HRRP enhancement preprocessing method according to claim 1, characterized in that: The step 8 comprises: Step 8.1, obtaining a first target spectrum signal; Step 8.2: obtaining a mask matrix and a position matrix of the first target spectrum signal according to the first target spectrum signal; Step 8.3, performing modulo-2 norm normalization processing on the amplitude of the first target spectrum signal to obtain a second target spectrum signal; Step 8.4: Pass the second target spectrum signal through the Embedding module, feature extraction module, context feature extraction module and interpolation module of the trained Transformer network model in sequence to obtain interpolation values ​​of missing values, and use the interpolation values ​​to fill the missing values ​​of the corresponding spectrum signal to obtain a reconstructed signal.

Images