A Radar Signal Interference Pattern Recognition Method and System Based on Transformer
Through the Transformer-based method, the problem of identification difficulty in traditional radar signal interference style recognition when the transmission waveform is variable is solved, and the effective identification of adaptive interference signals is achieved, which improves the recognition accuracy and inference speed.
Patent Information
- Application Number
- CN202410299393.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-03-15
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2044-03-15
AI Technical Summary
Traditional radar signal interference style recognition methods are difficult to adapt to adaptive interference signals when the transmission waveform is variable, resulting in poor recognition performance and the interference style under unseen parameters cannot be recognized.
Using a Transformer-based method, the timing radar signal segmentation, rearrangement, and 1×1 convolution feature extraction are used, and the deep feature enhancement is enhanced by using a multi-layer Transformer module. Combined with the deep perception machine and the SoftMax module to output the interference style probability, the maximum probability is selected. styles are identified.
The effective identification of different interference patterns is achieved under the condition of variable transmission waveforms, and under certain signal-to-noise ratio conditions, the recognition effect is comparable or better than the traditional CNN method, the model parameter amount is reduced by 50%, and the inference speed is increased by 20 times.
Smart Images

Figure CN118349823B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of pattern recognition and radar anti-interference, and specifically relates to a Transformer-based radar signal interference pattern recognition method and system, which are used to recognize the interference pattern to which the radar signal is subjected. Background Art
[0002] In cognitive radar systems, it is necessary to identify the interference patterns of radar signals and adaptively optimize the parameters of the transmission waveform based on the interference patterns. With the rapid development of interference technology, current jammers often use DRFM (Digital Radio Frequency Memory) interference technology to achieve deceptive interference by capturing, discontinuously sampling, modulating and forwarding radar signals. Due to the variability of the transmission waveform of cognitive radar, that is, the transmission waveform changes intelligently with the interference environment, the interference signal of the DRFM jammer also changes adaptively, resulting in a huge interference signal parameter space. The recognition of interference patterns under the condition of radar transmission waveform changes faces huge challenges. Traditional methods based on deep convolutional neural networks (CNN) can have good recognition effects when the transmission waveform is fixed or limited. When the interference signal has adaptive variability, even if the same interference pattern is selected differently, the characteristics of the interference signal, such as the characteristics obtained based on time-frequency processing, will also change. Due to the huge parameter space, it is difficult to exhaust the interference signals under each parameter combination and train the network. The traditional network has poor recognition performance and cannot recognize the interference pattern under zero samples (interference signals corresponding to no parameters). Summary of the invention
[0003] Purpose of the invention: In view of the problem that existing interference pattern recognition methods lack adaptability to transmission waveform parameters, the present invention proposes a Transformer-based radar signal interference pattern recognition method, which can be used to effectively identify various radar signal interference patterns when the transmission waveform is variable.
[0004] Another object of the present invention is to provide a radar signal interference pattern recognition system based on Transformer.
[0005] Technical solution: In the first aspect, a radar signal interference pattern recognition method based on Transformer includes the following steps:
[0006] The received time series radar signal is segmented and rearranged so that the time series signal becomes a time series vector with the same dimension;
[0007] Perform 1×1 convolution on the time series vector to perform preliminary feature extraction;
[0008] The time series vector after feature extraction is passed through D layers 1 The Transformer module performs deep feature enhancement and extraction;
[0009] The data after deep feature extraction is averaged to reduce the dimension;
[0010] The reduced data is passed through a depth D 2 After the deep perceptron composed of a fully connected layer and a PReLu activation function, the probability of each interference pattern is output through the SoftMax module, and the one with the largest probability is selected as the interference pattern.
[0011] Furthermore, each Transformer module includes a multi-head attention mechanism, residual connection, normalization, and fully connected layers. The multi-head attention mechanism divides the input vector of length K into H sub-vectors evenly, performs independent self-attention mechanism calculation on each sub-vector, and concatenates the output vectors to obtain the final output vector; the final output vector is subjected to residual connection, normalization, and fully connected layers to obtain feature-enhanced data. C is the number of channels of 1×1 convolution, L is the number of vectors after segmentation, and K is the segmentation length.
[0012] In the second aspect, a radar signal interference pattern recognition system based on Transformer is provided, including:
[0013] A segmentation and rearrangement module is used to segment and rearrange the received time series radar signal so that the time series signal becomes a time series vector with the same dimension;
[0014] The convolution module is used to perform 1×1 convolution on the time series vector for preliminary feature extraction;
[0015] Transformer module is used to transform the time series vector after feature extraction through D layers. 1 The Transformer module performs deep feature enhancement and extraction;
[0016] Dimensionality reduction module, used to reduce the dimension of the data after deep feature extraction by averaging;
[0017] The recognition module is used to transform the reduced-dimensional data through a depth D 2 After the deep perceptron composed of a fully connected layer and a PReLu activation function, the probability of each interference pattern is output through the SoftMax module, and the one with the largest probability is selected as the interference pattern.
[0018] According to a third aspect, a computer device is provided, comprising:
[0019] one or more processors;
[0020] Memory; and
[0021] One or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and when the programs are executed by the processors, the steps of the Transformer-based radar signal interference pattern recognition method as described in the first aspect of the present invention are implemented.
[0022] In a fourth aspect, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of the Transformer-based radar signal interference pattern recognition method as described above are implemented.
[0023] Beneficial effects: The present invention performs preliminary feature extraction on the received interfered radar signal by segmentation, rearrangement and 1×1 convolution preprocessing operations, and then outputs the probability of each interference pattern through feature enhancement and extraction of multi-layer Transformer modules, averaging dimension reduction, and a combination of deep perceptron and SoftMax, and finally selects the interference pattern with the highest probability. Simulation results show that the method of the present invention can identify different interference patterns, and achieves recognition effects comparable to or even better than CNN under certain signal-to-noise ratio conditions. Compared with existing implementation methods, by using a combination of time reshoot, 1×1 convolution, and Transformer modules instead of traditional CNN as a feature extractor, the model parameters are reduced to 1 / 50 and the model inference speed is increased by 20 times under the same recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] Figure 1 The present invention is a flowchart of a radar signal interference pattern recognition method based on Transformer;
[0025] Figure 2 A schematic diagram of the process flow of the Transformer module in the processing flow provided by the present invention. DETAILED DESCRIPTION
[0026] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.
[0027] The present invention is suitable for identifying interference patterns of radar signals. The overall process of the identification method is shown in the attached figure. Figure 1 As shown, the specific steps include:
[0028] Step 1) segmenting and rearranging the received time series radar signal so that the time series signal becomes a time series vector with the same dimension.
[0029] Assume that the received radar signal is s[n], each frame of radar signal has N sampling points, and the segmentation length is set to K, then each frame of signal can be divided into L = N / K continuous vectors. Rearrange these L vectors, then each frame of signal s[n] can be rearranged into
[0030] S r [l,:]=s[lK: (l+1)K], l=1, 2,...,L (1)
[0031] The rearranged signal S r It is composed of L consecutive K-dimensional vectors. The rearrangement enables the data to be used for subsequent convolution operations. Step 2) The rearranged time series vector s r Perform a 1×1 convolution.
[0032] For a kernel matrix S r is the matrix to be convolved, and the convolution operation can be calculated by the following formula:
[0033] S f [c, l, k] = (k * S r )[l, k] = ∑ p ∑ q S r (l+p,k+q)k(p,q) (2)
[0034] Here, the kernel matrix dimension m = 1, and its number of channels C can be set freely. The output feature matrix
[0035] Preprocessing and preliminary feature extraction can be completed by performing 1×1 convolution on the rearranged data.
[0036] Step 3) The initially extracted feature matrix is passed through a depth of D 1 The Transformer module performs deep feature enhancement and extraction.
[0037] For the input sequence For a single self-attention mechanism module, there is a parameter matrix Calculate the index, key, and value separately:
[0038] q i =W q x i k i =W k x i vi=W υ x i (3)
[0039] The weight calculation of each input can be obtained by the following formula:
[0040]
[0041] After obtaining the weights, the output vector sequence can be calculated step by step:
[0042] y i =∑ j w ij v j (5)
[0043] The input vector of length K is evenly divided into H sub-vectors, each sub-vector is independently calculated by the self-attention mechanism, and the output vectors are concatenated to obtain the final output vector.
[0044]
[0045] Through the above formula (6) and formulas (4) and (5), we can get the output result of the multi-head attention mechanism MultiHeadAttention
[0046] As attached Figure 2 A flow chart of the Transformer module is given, where the input is The data input to the Norm Layer is:
[0047]
[0048] NormLayer normalized the data:
[0049]
[0050] Among them, γ is the scaling factor, β is the offset factor, E is the mean, and Var is the variance.
[0051] As attached Figure 2 , FC represents the fully connected layer, PReLu is the activation function, and its expression is:
[0052]
[0053] Here, α is a hyperparameter.
[0054] The multi-head attention mechanism can be used as a deep feature extractor to extract deep features of time series data with very low parameter count and inference time. Residual connection and normalization make the model converge faster and avoid overfitting. Full connection as a post-processing part can correct the distribution of feature vectors. Figure 2 The calculation process can obtain the data after feature enhancement By using the Transformer module instead of the traditional convolutional layer for deep feature extraction, the number of model parameters can be greatly reduced and the model reasoning speed can be improved while achieving the same recognition accuracy.
[0055] Step 4) Reduce the dimension of the data after deep feature extraction by averaging.
[0056] Data Dimensionality reduction can be achieved by averaging:
[0057]
[0058] in,
[0059] By averaging and retaining global features instead of traditional forward propagation of all features, the number of model parameters can be greatly reduced and the inference speed can be improved.
[0060] Step 5) The interference pattern of the reduced-dimensional data is identified through a combination of a multi-layer perceptron and a SoftMax module, the probability of each interference pattern identification is output, and the interference pattern with the highest probability is selected as the interference pattern.
[0061] The multi-layer perceptron (MLP) consists of D layers. 2 The fully connected layer and PReLu activation function are composed of. The expression of PReLu activation function is shown in formula (9).
[0062] Then the probability of each interference pattern finally output is:
[0063] s p =SoftMax(MLP(S aυg )) (11)
[0064] The final recognition result is the maximum value among all probabilities, that is, argmax(s p ).
[0065] The above describes the implementation steps of the radar signal interference pattern recognition proposed by the present invention. In order to verify the effectiveness of this method, a simulation experiment is used to illustrate it. 1 Set to 2, D 2Set to 4, C to 1, K to 25, and H to 4. Each sample signal in the training set and the test set is an interfered signal generated by different combinations of eight different types of transmission signals and four different types of interference patterns. The transmission signal types include linear frequency modulation signals (LFM), two-phase coded signals (Barker codes), multi-phase coded signals (Frank codes, P1-P4 codes), and frequency modulation signals (Costas frequency modulation codes). The interference patterns include four main lobe interferences commonly used by DRFM jammers: intermittent sampling retransmission and forwarding interference (ISRJ), noise convolution interference (NCJ), distance deception interference (RDJ), and velocity deception interference (VDJ). The working process of a DRFM jammer usually consists of: reception, sampling, quantization, remodulation, and forwarding. Among them, quantization is usually phase quantization. If the received signal is y p (t) = p(t)e jφ(t) , then the quantized signal is:
[0066]
[0067] In the simulation, deception interference is to remodulate and retransmit the above quantized signal so that the signal received by the radar contains a large number of false targets.
[0068] The training set consists of 3200 signals, the test set consists of 800 signals, the number of sampling points of each signal is 6000, the signal-to-noise ratio is set to 8dB, and the interference-to-noise ratio is set to 20dB. During the training process, the number of batch samples is set to 64, the number of cycles is set to 50, the learning rate is set to 0.003, and the optimizer is selected as Adam. The accuracy, precision, recall rate, and F1 score of the invention for identifying interference patterns of various radar signals are shown in Table 1 after multiple simulation verifications.
[0069] Table 1 Values of evaluation indicators for each interference style
[0070]
[0071]
[0072] It can be seen that the present invention can identify interference patterns under a variety of transmission waveforms, and under simulation conditions, the average recognition accuracy is about 94%.
[0073] Furthermore, the simulation experiment provides the parameter quantity, memory usage, calculation amount and reasoning time of the present invention as shown in Table 2.
[0074] Table 2 Performance evaluation index values
[0075] Parameter quantity Memory usage Calculation Amount Reasoning time 15.63K 61.56K 3.70M 0.05s
[0076] It can be seen from Table 2 that the storage resources and computing resources occupied by the present invention are very limited, and the interference pattern recognition of radar signals can be performed very efficiently with good real-time performance.
[0077] According to another embodiment of the present invention, a radar signal interference pattern recognition system based on Transformer is provided, comprising:
[0078] A segmentation and rearrangement module is used to segment and rearrange the received time series radar signal so that the time series signal becomes a time series vector with the same dimension;
[0079] The convolution module is used to perform 1×1 convolution on the time series vector for preliminary feature extraction;
[0080] Transformer module is used to transform the time series vector after feature extraction through D layers. 1 The Transformer module performs deep feature enhancement and extraction;
[0081] Dimensionality reduction module, used to reduce the dimension of the data after deep feature extraction by averaging;
[0082] The recognition module is used to transform the reduced-dimensional data through a depth D 2 After the deep perceptron composed of a fully connected layer and a PReLu activation function, the probability of each interference pattern is output through the SoftMax module, and the one with the largest probability is selected as the interference pattern.
[0083] It should be understood that the Transformer-based radar signal interference pattern recognition system provided in this embodiment can implement all the technical solutions in the above-mentioned method embodiments, and the functions of its various functional modules can be specifically implemented according to the methods in the above-mentioned method embodiments. The specific implementation process can refer to the relevant description in the above-mentioned embodiments, which will not be repeated here.
[0084] The present invention also provides a computer device, comprising: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and when the programs are executed by the processors, the steps of the Transformer-based radar signal interference pattern recognition method as described above are implemented.
[0085] The present invention also provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the Transformer-based radar signal interference pattern recognition method as described above are implemented.
[0086] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, devices (systems), computer equipment or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.
[0087] The present invention is described with reference to a flowchart of a method according to an embodiment of the present invention. It should be understood that each process in the flowchart and a combination of processes in the flowchart can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the process in the flowchart. Figure 1 A device that specifies functions in a process or multiple processes.
[0088] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A function specified in a process or multiple processes.
[0089] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 The steps of a specified function in a process or multiple processes.
Claims
1. A radar signal interference pattern recognition method based on Transformer, characterized in that: The following steps are involved: The received time series radar signal is segmented and rearranged to make the time series signal become a time series vector with the same dimension, including: the received radar signal is s[n], each frame of radar signal has N sampling points, the segmentation length is set to K, each frame signal is divided into L = N / K continuous vectors, and these L vectors are rearranged, then each frame signal s[n] is rearranged into The rearranged signal S r It consists of L continuous K-dimensional vectors; Perform 1×1 convolution on the time series vector to perform preliminary feature extraction; The time series vector after feature extraction is passed through a Transformer module with a layer number of D1 for deep feature enhancement and extraction. Each Transformer module includes a multi-head attention mechanism, residual connection, normalization, and a fully connected layer. The multi-head attention mechanism divides the input vector of length K into H sub-vectors evenly, performs an independent self-attention mechanism calculation on each sub-vector, and concatenates the output vectors to obtain the final output vector; the final output vector is passed through a residual connection, normalization, and a fully connected layer to obtain the data after feature enhancement. C is the number of channels of 1×1 convolution, L is the number of vectors after segmentation, and K is the segmentation length; The data after deep feature extraction is averaged and reduced in dimension, expressed as: in, After the dimensionality reduction data passes through a deep perceptron with a depth of D2 consisting of a fully connected layer and a PReLu activation function, the probability of each interference pattern is output through the SoftMax module, and the one with the largest probability is selected as the interference pattern.
2. The method according to claim 1, characterized in that: When performing 1×1 convolution on the time series vector, the number of convolution channels C can be set freely, and the output feature matrix is 3. A radar signal interference pattern recognition system based on Transformer, characterized in that: include: The segmentation and rearrangement module is used to segment and rearrange the received time series radar signal so that the time series signal becomes a time series vector with the same dimension, including: the received radar signal is s[n], each frame of radar signal has N sampling points, the segmentation length is set to K, each frame signal is divided into L = N / K continuous vectors, and the L vectors are rearranged, so that each frame signal s[n] is rearranged into The rearranged signal S r It consists of L continuous K-dimensional vectors; The convolution module is used to perform 1×1 convolution on the time series vector for preliminary feature extraction; Transformer module is used to perform deep feature enhancement and extraction on the time series vector after feature extraction through a Transformer module with a layer number of D1. Each Transformer module includes a multi-head attention mechanism, residual connection, normalization, and a fully connected layer. The multi-head attention mechanism divides the input vector of length K into H sub-vectors evenly, performs independent self-attention mechanism calculation on each sub-vector, and concatenates the output vectors to obtain the final output vector; the final output vector is subjected to residual connection, normalization, and a fully connected layer to obtain the data after feature enhancement. C is the number of channels of 1×1 convolution, L is the number of vectors after segmentation, and K is the segmentation length; The dimension reduction module is used to reduce the dimension of the data after deep feature extraction by averaging, which is expressed as: in, The recognition module is used to pass the reduced-dimensional data through a deep perceptron with a depth of D2 consisting of a fully connected layer and a PReLu activation function, and then output the probability of each interference pattern through the SoftMax module, and select the one with the largest probability as the interference pattern.
4. A computer device, characterized in that: include: one or more processors; Memory; as well as One or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and when the programs are executed by the processors, the steps of the Transformer-based radar signal interference pattern recognition method as described in any one of claims 1 to 2 are implemented.
5. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the Transformer-based radar signal interference pattern recognition method as described in any one of claims 1 to 2 are implemented.
