An auxiliary data screening method based on a deep learning model

Abstract

The application provides an auxiliary data screening method based on a deep learning model, and relates to the technical field of radar signal processing, and comprises the following steps: constructing an auxiliary data screening model based on a Transformer architecture; obtaining radar echo data; assigning a label to each sample in the radar echo data, and constructing a binary classification data set; after the binary classification data set is preprocessed, the auxiliary data screening model is trained through a back propagation algorithm, and a trained auxiliary data screening model is obtained; obtaining to-be-detected unit data, inputting the to-be-detected unit data into the trained auxiliary data screening model after preprocessing, outputting the probability that each data contains an interference target, comparing the output probability with a pre-set probability threshold, and judging whether the to-be-detected unit data contains the interference target based on the comparison result. The auxiliary data screening model can automatically learn feature information in the radar echo data, is not affected by multiple interference targets, and can effectively distinguish whether the distance unit signal contains the interference target component.

Description

Technical Field

[0001] This application relates to the field of radar signal processing technology, and in particular to an auxiliary data filtering method based on a deep learning model. Background Technology

[0002] Space-time adaptive processing is a key technology for airborne radar to detect moving targets in strong clutter environments, and its performance depends on the clutter plus noise covariance matrix. Typically, the clutter plus noise covariance matrix is ​​unknown and needs to be estimated using auxiliary data from multiple range cells adjacent to the target cell, requiring these auxiliary data to have independent and identically distributed (ICD) characteristics. However, in practical applications, the auxiliary data may be contaminated by interfering targets, discrete clutter, etc., failing to meet the ICD assumption, leading to significant deviations in the covariance matrix estimation and thus affecting the detection performance of space-time adaptive processing.

[0003] To address the issue of auxiliary data contamination, researchers have proposed various non-uniform detection algorithms. Among them, the generalized inner product method is widely used in auxiliary data screening. It removes auxiliary data with clutter statistical characteristics different from the target cell by examining the covariance matrix. However, in the presence of multiple interfering targets, the examination covariance matrix inevitably contains components of interfering targets. Due to the "masking effect," the generalized inner product method cannot completely remove auxiliary data containing interfering targets. The interfering target components in the estimated clutter plus noise covariance matrix cause the target signal to self-cancel, thus significantly reducing the performance of spatiotemporal adaptive processing.

[0004] In recent years, deep learning has gradually attracted researchers' attention due to its powerful nonlinear modeling and feature extraction capabilities. How to introduce deep learning into data filtering tasks to address the impact of the "masking effect" and build more efficient and accurate filtering models urgently requires in-depth exploration and innovation. Summary of the Invention

[0005] This application aims to at least partially address one of the technical problems in the related art.

[0006] Therefore, the first objective of this application is to propose an auxiliary data filtering method based on a deep learning model. This method automatically learns the feature information in radar echo data through an auxiliary data filtering model based on the Transformer architecture, and is not affected by multiple interfering targets, effectively distinguishing whether the range cell signal contains interfering target components.

[0007] The second objective of this application is to propose an auxiliary data filtering system based on a deep learning model.

[0008] The third objective of this application is to propose a computer device.

[0009] To achieve the above objectives, the first aspect of this application proposes an auxiliary data filtering method based on a deep learning model, comprising:

[0010] Construct an auxiliary data filtering model based on the Transformer architecture;

[0011] Acquire radar echo data;

[0012] Assign a label to each sample in the radar echo data to construct a binary classification dataset;

[0013] After preprocessing the binary classification dataset, the auxiliary data filtering model is trained using the backpropagation algorithm to obtain the trained auxiliary data filtering model.

[0014] The system acquires data of the unit to be detected, preprocesses it, and then inputs it into the trained auxiliary data filtering model. It outputs the probability that each data point contains an interfering target, compares the output probability with a pre-set probability threshold, and determines whether the interfering target is contained based on the comparison result.

[0015] To achieve the above objectives, a second aspect of the present invention proposes a deep learning-based assisted data filtering system, comprising:

[0016] The model building module is used to build auxiliary data filtering models based on the Transformer architecture;

[0017] The data acquisition module is used to acquire radar echo data;

[0018] The dataset building module is used to assign labels to each sample in the radar echo data and build a binary classification dataset.

[0019] The model training module is used to preprocess the binary classification dataset and then train the auxiliary data filtering model through the backpropagation algorithm to obtain the trained auxiliary data filtering model.

[0020] The detection module is used to acquire the data of the unit to be detected, preprocess it and input it into the trained auxiliary data filtering model, output the probability that each data contains the interference target, compare the output probability with the pre-set probability threshold, and determine whether the interference target is contained based on the comparison result.

[0021] To achieve the above objectives, a third aspect of the present invention provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-described auxiliary data filtering method based on a deep learning model.

[0022] The auxiliary data filtering method based on a deep learning model in this application utilizes the superiority of the Transformer architecture model in processing sequential information. Combined with the spatiotemporal two-dimensional structure of radar echo data, an auxiliary data filtering model is constructed. This model fully explores the differences between radar echoes containing and not containing interfering target components. It can determine whether the unit to be detected contains interfering targets without relying on echo data from other range gates or prior information on the number of interfering targets. This method specifically solves the problems of the generalized inner product method, such as susceptibility to the "masking effect" in scenarios with dense interfering targets and performance degradation due to the unknown number of interfering targets. It improves the removal rate of auxiliary data containing interfering target components.

[0023] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0024] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0025] Figure 1 This is a flowchart illustrating an auxiliary data filtering method based on a deep learning model provided in Embodiment 1 of this application.

[0026] Figure 2 This is a network structure diagram of the auxiliary data filtering model in an embodiment of this application;

[0027] Figure 3 The figure shows a performance comparison between the embodiments of this application and the traditional generalized inner product method and the iterative elimination generalized inner product method;

[0028] Figure 4 This is a schematic diagram of the structure of an auxiliary data filtering system based on a deep learning model, provided in an embodiment of this application. Detailed Implementation

[0029] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0030] The following description, with reference to the accompanying drawings, illustrates an embodiment of an assisted data filtering method and system based on a deep learning model.

[0031] Figure 1 This is a flowchart illustrating an auxiliary data filtering method based on a deep learning model, as provided in Embodiment 1 of this application.

[0032] like Figure 1 As shown, this deep learning model-based assisted data filtering method includes the following steps:

[0033] Step 101: Construct an auxiliary data filtering model based on the Transformer architecture;

[0034] In this embodiment, the auxiliary data filtering model based on the Transformer architecture includes an input embedding layer, a positional encoding layer, a multi-head attention layer, a feedforward network layer, a normalization layer, and an output classification layer, specifically:

[0035] Input Embedding Layer: Receives radar echo data as input data. It has a two-dimensional space-time structure. Each radar echo data consists of multiple pulses and multiple receiving channels, corresponding to the time dimension and the spatial dimension, respectively.

[0036] The input data sequence length is equal to the number of pulses, and each sequence element is a multi-dimensional vector whose dimension is equal to the number of receiving channels. Embedding feature processing is performed on the input data through a non-linear mapping, mapping all receiving channel data of each pulse into a high-dimensional embedding vector, thus obtaining the embedding feature representation of each pulse, expressed as:

[0037] InputEmdedding(X in ) = X e

[0038] in, To input radar echo data, Let X be the embedding feature, N be the number of receiving channels, M be the number of pulses, and d be the embedding dimension. e The data from the N receiving channels representing M pulses are each mapped to an embedding feature of dimension d;

[0039] Positional encoding layer: Employs an encoding method based on sine and cosine function generation to embed feature X. e Add pulse sequence encoding information, the expression is:

[0040] PositionalEncoding(X e ) = X e +P=X e

[0041] in, For location information, The encoding method expression for the sine and cosine functions generated for the embedded features after adding location information is as follows:

[0042] P pos,2i =sin(pos / 10000) 2i / d )

[0043] Ppos,2i+1 =cos(pos / 10000) 2i / d )

[0044] Where pos represents the pulse sequence;

[0045] Multi-head attention layer: Based on a scaled dot product self-attention mechanism, multiple parallel attention heads calculate the attention distribution for the embedded features of each pulse, and then perform weighted fusion to generate the corresponding feature representation. Assume the multi-head attention mechanism layer receives a matrix... The query matrix is ​​obtained after three different linear transformations. Key matrix Value matrix The calculation formula is:

[0046] Q = W q ·X mha +b q

[0047] K = W k ·X mha +b k

[0048] V = W v ·X mha +b v

[0049] in, and Let h be the parameter matrix for the linear transformation. The multi-head attention mechanism divides the query matrix, key matrix, and value matrix into h heads. i Let represent the i-th head in multi-head attention, and correspondingly, the query matrix for each head is . The key matrix is The value matrix is Where Q = {Q1, Q2, ..., Q} h}, K = {K1, K2, ..., K} h}, V={V1,V2,...,V h Each head performs a scaled dot product attention operation once, expressed as:

[0050]

[0051] Here, Attention represents the scaled dot product attention operation, and softmax represents the activation function. In the attention mechanism, Q... i and K i The correlation between the embedded features of different pulses is calculated by performing a dot product operation, and then the attention weight distribution is obtained by using the softmax function. Used for weighting matrix Vi As each head i The outputs of all heads are concatenated and subjected to a linear transformation to obtain the final output of the multi-head attention layer, expressed as:

[0052] MHA(x mha ) = W l ·concat(head1,head2,...,head h )+b l

[0053] Here, concat represents the concatenation operation. and Let be the parameter matrix of the linear transformation;

[0054] Feedforward network layer: Composed of a multilayer perceptron, it performs a nonlinear transformation on the embedded features of each pulse to further mine potential higher-order features. Assume the feedforward network layer's receiver matrix... The calculation formula for this layer is:

[0055]

[0056] in, and Let d be the parameter matrix of the linear transformation. ′ is the number of neurons in the first hidden layer of the multilayer perceptron, and f represents the ReLU nonlinear activation function;

[0057] Normalized layer: The layer normalization method is used, and the calculation formula is as follows:

[0058]

[0059] Where x is the input data, E[x] and Var[x] represent the mean and variance of the input data, ∈ is a very small constant used to avoid division by zero, and γ and β are learnable parameters.

[0060] Output classification layer: Consists of a fully connected network that performs the final discrimination based on the global features of the radar echo data. It maps the global features to the output space through a linear transformation and combines this with a softmax function to output the probability value of each input data point containing a jamming target. Assume the final output of the network is... The expression for the output classification layer is:

[0061] FCN(X o ) = w o ·flatten(X o )+b o

[0062] Here, "flatten" indicates the flattening operation. and Here, n is the number of classes. In the auxiliary data filtering, it is necessary to classify samples containing interfering targets and samples not containing interfering targets, so n=2. Finally, the output of the classification layer is converted into class probabilities through the softmax function to obtain the probability of containing interfering targets.

[0063] In this embodiment, according to Figure 2 An auxiliary data filtering model was constructed. The model's network structure consists of an input embedding layer, a position encoding layer, followed by a multi-head attention layer, a normalization layer, a feedforward network layer, and another normalization layer stacked 8 times. Residual connections were added between the input and output of the multi-head attention layer and between the input and output of the feedforward network layer. The final layer of the model's network structure is the output classification layer. The model receives input radar echo data with dimensions N=11 and M=32. The embedding feature dimension for each pulse is d=64. The number of heads in the multi-head attention mechanism layer is h=8, and the number of neurons in the first hidden layer of the multilayer perceptron is d. ′ =128.

[0064] Step 102: Acquire radar echo data;

[0065] In this embodiment, the radar echo data refers to the echo data of M pulses from N array elements acquired by the radar antenna array within one coherent pulse time interval. The expression is:

[0066]

[0067] Where H0 represents the case without interfering targets, H1 represents the case with interfering targets, and x c For clutter signals, x t The target signal is the interference signal, and n is the noise signal.

[0068] Considering the impact of array element channel amplitude and phase errors on radar received signals, clutter signal x c The expression is:

[0069]

[0070] Where, N c It is the number of clutter blocks within a distance ring, α i , and f D,i Let f represent the complex amplitude, spatial cone angle, and normalized Doppler frequency of the i-th clutter block, respectively. D,i =2vcosψ i / (λf r), where v is the speed of the carrier platform, λ is the transmission wavelength, and f r Let e ​​be the pulse repetition frequency, and e = [e1, e2, ..., e]. N [ ] represents the amplitude and phase error vector of the array element channel. and Let S(ψ) represent the amplitude error and phase error corresponding to the r-th array element, respectively, where r = 1, 2, ..., N, and ⊙ denotes the Hadamard product. i ,f D,i S represents the space-time steering vector of the i-th clutter block. D,i and S s,i These are the time steering vector and spatial steering vector of the i-th clutter block, respectively, and their specific expressions are as follows:

[0071]

[0072] Where, d a This refers to the element spacing of the radar antenna array.

[0073] Considering the impact of array element channel amplitude and phase errors on radar received signals, the jamming target signal x t The expression is:

[0074]

[0075] Where, β t To interfere with the complex amplitude of the target, S(f) D,t ,f s,t S is the space-time steering vector of the jamming target. D,t and S s,t These are the time steering vector and spatial steering vector of the interference target, respectively, and their specific expressions are as follows:

[0076]

[0077]

[0078] Among them, f D,t and f s,t These are the normalized Doppler frequency and spatial frequency of the interfering target, respectively.

[0079] The noise signal n is zero-mean complex Gaussian white noise.

[0080] Step 103: Assign a label to each sample in the radar echo data to construct a binary classification dataset;

[0081] In this embodiment, radar echo data containing only clutter and noise components are considered as samples without interfering targets, and the label is set to 0. For radar echo data containing clutter, interfering targets, and noise components, the label is set to 1. The label information is used to calculate the loss function and backpropagate to update the model parameters during the training process.

[0082] In this embodiment, the clutter and noise components in samples without interfering targets and samples with interfering targets are completely identical. The only difference is the presence or absence of interfering target components. This ensures that the samples have representative differences under the condition of consistent background components, allowing the model to focus more on learning whether there are interfering target components and improve the model's feature generalization ability.

[0083] In this embodiment, the number of samples without interfering targets and the number of samples with interfering targets are the same to avoid sample imbalance.

[0084] Step 104: After preprocessing the binary classification dataset, train the auxiliary data filtering model using the backpropagation algorithm to obtain the trained auxiliary data filtering model.

[0085] In this embodiment, since radar echo data is complex data and the model network only accepts real input data, a modulo operation is performed on the radar echo data to convert the complex data into real data. Then, a maximum-minimum normalization operation is performed, and the calculation expression is:

[0086]

[0087] Here, min represents taking the minimum value, and max represents taking the maximum value. The preprocessed data has a dimension of NM×1, which is then converted to N×M to match the network input dimension requirements.

[0088] Preferably, the batch size is set to 4096, the learning rate is set to 0.0001, the cross-entropy loss is used as the loss function, the Adam optimizer is used for iterative training, and the network parameters of the auxiliary data filtering model are optimized through the backpropagation algorithm.

[0089] Step 105: Obtain the data of the unit to be detected, preprocess it and input it into the trained auxiliary data filtering model, output the probability that each data contains the interference target, compare the output probability with the preset probability threshold, and determine whether the interference target is contained based on the comparison result.

[0090] In this embodiment, the data of the unit to be detected is sequentially processed by modulo, maximum-minimum normalization and dimensionality transformation, and then input into the trained auxiliary data filtering model. The model outputs the probability that the current data of the unit to be detected contains an interfering target. The output probability is compared with a pre-set probability threshold, and the model determines whether it contains an interfering target based on the comparison result.

[0091] Preferably, the probability threshold is set to 0.5.

[0092] The auxiliary data filtering method based on a deep learning model in this application utilizes the superiority of the Transformer architecture model in processing sequential information. Combined with the spatiotemporal two-dimensional structure of radar echo data, an auxiliary data filtering model is constructed. This model fully explores the differences between radar echoes containing and not containing interfering target components. It can determine whether the unit to be detected contains interfering targets without relying on echo data from other range gates or prior information on the number of interfering targets. This method specifically solves the problems of traditional generalized inner product methods, such as susceptibility to the "masking effect" in scenarios with dense interfering targets and performance degradation due to the unknown number of interfering targets. It improves the removal rate of auxiliary data contaminated by interfering targets.

[0093] This paper compares the performance of the deep learning-based auxiliary data filtering method proposed in this application with that of the traditional Generalized Inner Product (GIP) and Reiterative Censored Generalized Inner Product (RCGIP) methods for auxiliary data filtering based on KASSPER data. KASSPER is a high-fidelity site-specific radar dataset based on heterogeneous terrain in a certain region. It is simulated using the SCATS (Splatter, Clutter, and Target Signal) model and includes many real-world effects, including heterogeneous terrain, subspace leakage, array error, and ground traffic. The KASSPER dataset includes 32 pulses, 11 receiving channels, and 1000 range gates. Considering the non-uniformity of the data and computational complexity, the traditional GIP and Reiterative Censored Generalized Inner Product methods only use the 32 pulses from the first receiving channel and divide the range cells into 10 segments for segmented algorithm implementation. Five interfering targets with equal power and normalized Doppler frequencies distributed in the interval [0.1, 0.2] are added to each segment. Their range gates are randomized, meaning that 5 out of every 100 range gates contain interfering targets, resulting in a total of 50 interfering targets to be eliminated. The elimination probability is defined as the percentage of correctly eliminated interfering targets out of the total number of interfering targets. Performance is evaluated through a Monte Carlo experiment with 1000 independent trials. The performance comparison results of this embodiment with the traditional generalized inner product method and the iterative elimination generalized inner product method are as follows: Figure 3 As shown, it can be seen that as the power of the interfering target increases, the elimination probability of the embodiment of this application gradually increases, and it is always better than the traditional generalized inner product method and the iterative elimination generalized inner product method. When the power of the added interfering target is 29dB or above, the embodiment of this application almost eliminates all range gates containing interfering targets, indicating that the embodiment of this application can effectively eliminate range gates containing interfering targets without being affected by multiple interfering targets.

[0094] To implement the above embodiments, this application also proposes an auxiliary data filtering system based on a deep learning model.

[0095] Figure 4 This is a schematic diagram of the structure of an auxiliary data filtering system based on a deep learning model, provided in an embodiment of this application.

[0096] like Figure 4 As shown, this deep learning model-based assisted data filtering system includes:

[0097] The model building module is used to build auxiliary data filtering models based on the Transformer architecture;

[0098] The data acquisition module is used to acquire radar echo data;

[0099] The dataset building module is used to assign labels to each sample in the radar echo data and build a binary classification dataset.

[0100] The model training module is used to preprocess the binary classification dataset and then train the auxiliary data filtering model through the backpropagation algorithm to obtain the trained auxiliary data filtering model.

[0101] The detection module is used to acquire the data of the unit to be detected, preprocess it and input it into the trained auxiliary data filtering model, output the probability that each data contains the interference target, compare the output probability with the pre-set probability threshold, and determine whether the interference target is contained based on the comparison result.

[0102] It should be noted that the foregoing explanation of the embodiment of the deep learning model-based auxiliary data filtering method also applies to the deep learning model-based auxiliary data filtering system of this embodiment, and will not be repeated here.

[0103] To implement the above embodiments, the present invention also proposes a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the methods described in the above embodiments.

[0104] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0105] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0106] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0107] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0108] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0109] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0110] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0111] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A data filtering method based on a deep learning model, characterized in that, include: Construct an auxiliary data filtering model based on the Transformer architecture; Acquire radar echo data; Assign a label to each sample in the radar echo data to construct a binary classification dataset; After preprocessing the binary classification dataset, the auxiliary data filtering model is trained using the backpropagation algorithm to obtain the trained auxiliary data filtering model. Data of the unit to be detected is acquired, preprocessed, and then input into the trained auxiliary data filtering model. The probability that each data point contains an interfering target is output. The output probability is compared with a pre-set probability threshold, and the interfering target is determined based on the comparison result. The auxiliary data filtering model based on the Transformer architecture includes an input embedding layer, a positional encoding layer, a multi-head attention layer, a feedforward network layer, a normalization layer, and an output classification layer. The input embedding layer receives radar echo data as input data. The radar echo data has a two-dimensional space-time structure. Each radar echo data consists of multiple pulses and multiple receiving channels, corresponding to the time dimension and the spatial dimension, respectively. The input data sequence length is equal to the number of pulses, and each sequence element is a multi-dimensional vector whose dimension is equal to the number of receiving channels. Embedding feature processing is performed on the input data through a non-linear mapping, mapping all receiving channel data of each pulse into a high-dimensional embedding vector, thus obtaining the embedding feature representation of each pulse, expressed as: in, To input radar echo data, For embedding features, Number of receive channels The number of pulses. For the embedding dimension, the embedded features represent pulse The data from each receiving channel is mapped to a dimension of... Embedded features; The position encoding layer employs an encoding method based on sine and cosine functions to embed features. Add pulse sequence encoding information, the expression is: in, For location information, The encoding method expression for the sine and cosine functions generated for the embedded features after adding location information is as follows: in, Indicates the order of pulses; The multi-head attention layer, based on a scaled dot product self-attention mechanism, calculates the attention distribution for the embedded features of each pulse through multiple parallel attention heads, and performs weighted fusion to generate the corresponding feature representation. Assuming a multi-head attention mechanism layer receiving matrix The query matrix is ​​obtained through three different linear transformations. Key matrix Value matrix The calculation formula is: in, and Let be the parameter matrix of the linear transformation. The multi-head attention mechanism divides the query matrix, key matrix, and value matrix into three parts. Size, The first in the bullish attention There are 1 head, and the query matrix corresponding to each head is . The key matrix is The value matrix is ,in, , , Each head performs a scaled dot product attention operation once, expressed as: in, This indicates the scaling dot product attention operation. This represents the activation function. In attention mechanisms, and The correlation between the embedding features of different pulses is calculated by performing a dot product operation, and then... The function yields the attention weight distribution. Used for weighted matrices As each head The outputs of all heads are concatenated and subjected to a linear transformation to obtain the final output of the multi-head attention layer, expressed as: in, This indicates a splicing operation. and Let be the parameter matrix of the linear transformation; The feedforward network layer consists of a multilayer perceptron. The multilayer perceptron performs a nonlinear transformation on the embedded features of each pulse to further mine potential higher-order features. Assuming the feedforward network layer receive matrix The calculation formula for this layer is: in, , and , Let be the parameter matrix of the linear transformation. This represents the number of neurons in the first hidden layer of a multilayer perceptron. Represents the ReLU nonlinear activation function; The normalization layer employs a layer normalization method, and the calculation formula is as follows: in, For input data, and This represents the mean and variance of the input data. It is a very small constant used to avoid division by zero. and These are learnable parameters; The output classification layer, consisting of a fully connected network, performs the final discrimination of the global features of the radar echo data. It maps the global features to the output space through a linear transformation and combines this with... The function outputs the probability value that each input data point contains the interfering target. Assume the final output of the network is The expression for the output classification layer is: in, Indicates the flattening operation. and For learnable parameters, To determine the number of categories, in the auxiliary data filtering process, samples containing interfering targets and those not containing interfering targets are classified. The output of the classification layer is passed through The function is converted into class probabilities, yielding the probability of the target containing interference.

2. The method as described in claim 1, characterized in that, The construction of a binary classification dataset using the radar echo data includes: Radar echo data containing clutter and noise components are treated as samples without interfering targets and labeled with 0. Radar echo data containing clutter, interfering targets, and noise components are treated as samples with interfering targets and labeled with 1.

3. The method as described in claim 1, characterized in that, The clutter and noise components in the samples without interfering targets and the samples containing interfering targets are completely identical; the difference lies in the presence or absence of interfering target components.

4. The method as described in claim 1, characterized in that, After preprocessing the binary classification dataset, the auxiliary data filtering model is trained using the backpropagation algorithm to obtain the trained auxiliary data filtering model, including: Modulo the complex data in the binary classification dataset to convert the complex data into real data, perform maximum and minimum value normalization, scale the data scale to between 0 and 1, and perform dimension transformation to adapt the data format to the input requirements of the model. The auxiliary data filtering model is trained using a preprocessed binary classification dataset. During training, cross-entropy loss is used as the loss function, and the Adam optimizer is used for iterative training. The network parameters of the model are optimized through the backpropagation algorithm.

5. The method as described in claim 1, characterized in that, The step of comparing the output probability with a pre-set probability threshold and determining whether it contains interfering targets based on the comparison result includes: If the output probability is less than the probability threshold, it is determined that there is no interfering target; If the output probability is not less than the probability threshold, it is judged to be a target containing interference.

6. A deep learning-based assisted data filtering system, characterized in that, The system implements the method as described in claim 1, the system comprising: The model building module is used to build auxiliary data filtering models based on the Transformer architecture; The data acquisition module is used to acquire radar echo data; A dataset construction module is used to assign labels to each sample in the radar echo data and construct a binary classification dataset. The model training module is used to preprocess the binary classification dataset and then train the auxiliary data filtering model through the backpropagation algorithm to obtain the trained auxiliary data filtering model. The detection module is used to acquire the data of the unit to be detected, preprocess it and input it into the trained auxiliary data filtering model, output the probability that each data contains an interfering target, compare the output probability with a pre-set probability threshold, and determine whether it contains an interfering target based on the comparison result.

7. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the assisted data filtering method based on a deep learning model as described in any one of claims 1-5.

Images