A radar intermittent sampling and forwarding interference detection method based on decision tree

Through the decision tree-based detection method, combined with the Transformer network and CFAR detector, the problem of insufficient detection performance under low interference signal-to-noise ratio is solved, and robust detection of intermittent sampling and forwarding interference of radar is achieved, adapting to periodicity and aperiodicity Interference, good mobility and detection accuracy.

CN118709083BActive Publication Date: 2024-11-15CHINA UNIV OF PETROLEUM (EAST CHINA)
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411195674.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-29
Publication Date
2024-11-15
Estimated Expiration
2044-08-29

AI Technical Summary

Technical Problem

The prior art has weak detection performance under low interference signal-to-noise ratio conditions and cannot adapt to periodic and non-periodic radar intermittent sampling forwarding interference forms.

Method used

A decision tree-based detection method is adopted, combined with Transformer network and CFAR detector, and through standardized processing, multi-head attention mechanism and decision tree integration, robust detection of interfering signals is achieved.

Benefits of technology

It has robust detection performance under low interference signal-to-noise ratio conditions, can adapt to dense and non-periodic interference, provide good mobility and detection accuracy, and is close to the lower limit of performance of traditional CFAR detectors.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118709083B_ABST
    Figure CN118709083B_ABST
Patent Text Reader

Abstract

The present invention provides a radar intermittent sampling forwarding interference detection method based on a decision tree, which relates to the technical field of radar interference detection, and specifically includes the following steps: standardizing radar signals; using a Transformer network to detect interference signals and obtain the detection results of the Transformer network; using the detection results of the Transformer as a reference, distinguishing interference signals from non-interference signals, calculating the detection threshold, and using a CFAR detector to further detect interference signals; using a decision tree to integrate the detection results obtained through steps S2 and S3, the instantaneous power and average power of the signal, and adopting different result selection strategies under different interference powers to obtain the final detection results. The technical solution of the present invention overcomes the problems in the prior art that the detection performance is weak under low interference signal-to-noise ratio conditions and cannot adapt to periodic and non-periodic interference forms.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of radar interference detection, and in particular to a radar intermittent sampling and forwarding interference detection method based on a decision tree. Background Art

[0002] Electronic warfare plays an increasingly important role in modern warfare. It uses electronic countermeasures (ECCM) technology to generate high-power active interference to suppress the normal detection functions of electronic systems such as radar, communication and reconnaissance. Intermittent sampling and forwarding jamming (ISRJ) is a typical and effective method to suppress the normal operation of radar in actual electronic warfare. ISRJ has a large dynamic range of power, non-periodic transmission and complex modulation characteristics, which brings great difficulties to interference suppression. In recent years, there have been many studies on ISRJ suppression, aiming to achieve interference filtering and successfully realize real target detection. The premise of interference suppression is to accurately locate the interference area in the echo. However, ISRJ has the characteristics of dense distribution, wide dynamic power range and non-periodic transmission, which makes it challenging to accurately detect the interference area in practice.

[0003] In the prior art, interference area detection methods are mainly divided into three categories: constant false alarm rate (CFAR) detection methods, feature domain detection methods, and machine learning-based methods. The traditional CFAR method only focuses on the power difference between the interference and pure target signals, and cannot utilize the characteristics of the continuous and dense distribution of interference, which significantly limits the detection performance under low interference signal-to-noise ratio (JSNR) conditions. In addition to the power difference, the feature domain detection method considers interference features in other domains, such as the time-frequency domain, semi-parametric domain, and probability and statistical domain, but the performance will degrade when the ISRJ is modulated, and the detection performance will be significantly affected under low JSNR conditions. Machine learning also faces the problems of transferability and interpretability.

[0004] Therefore, there is a need for a radar intermittent sampling and forwarding interference detection method based on a decision tree that has robust detection performance under low interference signal-to-noise ratio conditions and can adapt to periodic and non-periodic interference forms. Summary of the invention

[0005] The main purpose of the present invention is to provide a radar intermittent sampling forwarding interference detection method based on a decision tree to solve the problems in the prior art of weak detection performance under low interference signal-to-noise ratio conditions and inability to adapt to periodic and non-periodic interference forms.

[0006] To achieve the above object, the present invention provides a radar intermittent sampling and forwarding interference detection method based on a decision tree, which specifically includes the following steps:

[0007] S1, standardize the radar signal, including modulus normalization and length clipping.

[0008] S2, use the Transformer network to detect the interference signal and obtain the detection result of the Transformer network.

[0009] S3, using the detection results of Transformer as a reference, distinguishes interference signals from non-interference signals, calculates the detection threshold, and uses the CFAR detector to further detect interference signals.

[0010] S4, using a decision tree to integrate the detection results obtained in steps S2 and S3, the instantaneous power and the average power of the signal, and adopting different result selection strategies under different interference powers to obtain the final detection result.

[0011] Furthermore, the Transformer network in step S2 includes: an encoder layer, a decoder layer and a linear output layer connected to each other, wherein the encoder layer is composed of a plurality of encoder blocks, and the decoder layer is composed of a plurality of decoder blocks.

[0012] Furthermore, the signal data in step S2 is specifically processed as follows at the encoder layer:

[0013] S2.1.1, the signal data with a length of 1024 is added to the position encoder after passing through the linear embedding layer, where the calculation formula of the position encoder in the Transformer is as follows:

[0014] (1);

[0015] (2);

[0016] in, and Represent the codes of the signal at even and odd positions respectively. Indicates the position of a signal data in the entire signal data. express The dimension of Represents the position of a vector.

[0017] S2.1.2, the signal data then enters the encoder block and passes through the multi-head attention mechanism in the encoder block As well as residual connections and layer normalization, the multi-head attention mechanism for:

[0018] (3);

[0019] In the formula, represent The number of columns in the matrix, i.e. the vector dimension, Indicates that each row of the matrix is ​​normalized, where is the query matrix, is the key matrix, is a value matrix.

[0020] S2.1.3, after the forward neural network layer, residual connection and layer normalization, the output of an encoder block is obtained. The output of the previous encoder block is the input of the next encoder block, and the output of the last encoder block is the final output of the encoder layer.

[0021] Furthermore, the signal data in step S2 is specifically processed as follows at the decoder layer:

[0022] S2.2.1, same as step S2.1.1, the signal data with a length of 1024 is added to the position encoder after passing through the linear embedding layer.

[0023] S2.2.2, the signal data then passes through the first multi-head attention mechanism in the decoder block. The first multi-head attention mechanism uses a mask operation, and then undergoes residual connection and layer normalization.

[0024] S2.2.3, then passes through the second multi-head attention mechanism. The K and V matrices in the second multi-head attention mechanism are calculated using the output of the encoder layer, while Q is calculated using the output of the previous decoder block.

[0025] S2.2.4, after the forward neural network layer, residual connection and layer normalization, the output of a decoder block is obtained. The output of the previous decoder block is the input of the next decoder block, and the output of the last decoder block is the final output of the decoder layer.

[0026] S2.2.5, the final output of the decoder layer is obtained after the linear layer and the Softmax layer. The output result is in the form of marking the interference position as '1' and the non-interference position as '0'.

[0027] Furthermore, step S3 specifically includes the following steps:

[0028] S3.1, when there is no interference signal input Assumption: When the input signal includes interference signal, radar signal and noise signal, Assume that, expressed as formula (4):

[0029] (4);

[0030] in, is the noise signal, is the radar signal, is an interference signal.

[0031] S3.2, judgment threshold voltage The calculation formula is as follows:

[0032] (5);

[0033] in, is the false alarm rate, is the average power of the non-interference signal, calculated as follows:

[0034] (6);

[0035] In the formula Indicates the signal The amplitude of a signal, is the signal length of the non-interference position in a signal obtained according to the Transformer detection result.

[0036] S3.3, calculated by formula (5) , , Decision threshold voltage under three false alarm rate values , when the voltage value of the detection signal is greater than When it is marked as '1', it means there is interference signal, and when it is less than The time mark '0' represents no interference signal.

[0037] Furthermore, step S4 specifically includes the following steps:

[0038] S4.1, use the classification and regression tree CART algorithm to create a decision tree. The calculation formula of the Gini coefficient is as follows:

[0039] (7);

[0040] In the formula, represents the number of categories, Represents the probability that a sample belongs to this category, Represents the dataset The probability that two randomly selected samples will have different categories.

[0041] S4.2, for the dataset , according to the characteristics Can you take a value? , divide the sample into two data sets and , satisfying the following relationship:

[0042] (8).

[0043] Furthermore, step S4 further includes the following steps:

[0044] S4.3, in the characteristics Under the condition of The Gini coefficient Defined as:

[0045] (9);

[0046] in, , , Represents the data set , and The number of .

[0047] S4.4, select: average power ratio, instantaneous power ratio, Transformer detection results, and CFAR detector false alarm rate , , The three detection results at the time of are used as reference features of the decision tree, and the average power ratio is calculated as follows:

[0048] (10);

[0049] In the formula, It represents the average power of the interference signal in a signal segment. It represents the average power of non-interference signal in a signal. Indicates the average power ratio.

[0050] The instantaneous power ratio is calculated as follows:

[0051] (11);

[0052] In the formula Indicates the number 1 in a signal The amplitude of a signal, It represents the average power of non-interference signal in a signal. Indicates the number of signal points contained in a signal. Indicates The instantaneous power ratio of the signals.

[0053] The present invention has the following beneficial effects:

[0054] (1) The present invention has robust detection performance under low interference signal-to-noise ratio (JSNR) conditions. The continuous distribution features of interference are extracted by Transformer using position encoding and multi-head attention mechanism. This is an effective feature for judging interference signals under low JSNR conditions because the continuous distribution features bring information accumulation to weak interference signals.

[0055] (2) The present invention has the robust detection capability of dense interference. It adopts a hierarchical detection framework to divide the interference detection task into several sub-detection problems, effectively solving the difficulty of dense interference detection.

[0056] (3) The present invention has good mobility and can adapt to periodic and non-periodic interference forms. The method combines data priors and probability statistical models, and the CFAR detector provides the present invention with a performance lower limit close to that of the traditional CFAR detector. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the specific implementation of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the specific implementation or the prior art description. Obviously, the drawings described below are some implementations of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. In the drawings:

[0058] Figure 1 The loss curve is shown when the Transformer network is trained 100 times.

[0059] Figure 2 The Transformer network detection results are shown when JSNR is 6dB.

[0060] Figure 3 The Transformer network detection results at JSNR of 15 dB are shown.

[0061] Figure 4 The Transformer network detection results at JSNR of 30 dB are shown.

[0062] Figure 5 A statistical graph of Transformer detection probability is shown.

[0063] Figure 6 The CFAR detection results are shown when the JSNR is 6dB.

[0064] Figure 7 The CFAR detection results are shown when the JSNR is 15dB.

[0065] Figure 8 The CFAR detection results are shown when the JSNR is 30dB.

[0066] Fig. 9 A statistical graph of CFAR detection probability is shown.

[0067] Fig.10 A partial decision tree diagram is shown.

[0068] Fig.11 A sample of decision tree detection results and detection probability at JSNR 6 dB are shown.

[0069] Fig.12 A sample of decision tree detection results and detection probability at JSNR 15 dB are shown.

[0070] Fig.13 A sample of decision tree detection results and detection probability at JSNR 30 dB are shown.

[0071] Fig.14 The graphs showing the changes of decision tree detection probability, Transformer detection probability and CFAR detection probability with JSNR under ISDJ are shown. DETAILED DESCRIPTION

[0072] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0073] A radar intermittent sampling and forwarding interference detection method based on a decision tree specifically comprises the following steps:

[0074] S1, standardize the radar signal, including modulus normalization and length clipping.

[0075] S2, use the Transformer network to detect the interference signal and obtain the detection result of the Transformer network.

[0076] S3, using the detection results of Transformer as a reference, distinguishes interference signals from non-interference signals, calculates the detection threshold, and uses the CFAR detector to further detect interference signals.

[0077] S4, using a decision tree to integrate the detection results obtained in steps S2 and S3, the instantaneous power and the average power of the signal, and adopting different result selection strategies under different interference powers to obtain the final detection result.

[0078] The present invention first performs standardization processing (modulus normalization, length clipping, etc.) on the radar signal for subsequent processing. Then, the Transformer network is used to detect the radar intermittent sampling and forwarding interference signals, and the detection results of the Transformer network are obtained. The position where interference is detected is marked as '1', and the position without interference is marked as '0', generating a detection result sequence with the same length as the detection signal. This detection method can accurately detect the position of most interference signals, but detection errors may occur at the junction of interference signals and non-interference signals. In order to minimize this error, the following operations will be performed. In the second part, the CFAR detector is used to detect the interference signal, and the interference signal and non-interference signal are distinguished with the detection result of the Transformer as a reference. The non-interference signal is used to estimate the correct detection threshold, and then the CFAR is used to detect the interference signal to obtain the CFAR detection result. Under different interference powers, Transformer detection and CFAR detection have their own advantages and disadvantages. The present invention hopes that the final result can show the advantages of Transformer detection and CFAR detection and eliminate their disadvantages. Finally, a decision tree is used to integrate the detection results obtained by the two detection methods, as well as the instantaneous power and average power of the signal. Different result selection strategies are adopted under different interference powers to obtain the final detection result.

[0079] At present, the mainstream radar intermittent sampling and forwarding interference is mostly dense and continuous slice interference. The power of radar intermittent sampling and forwarding interference will be higher than the power of radar signal, showing that the amplitude of interference signal is higher than that of radar signal. Based on this phenomenon, interference signal and radar signal can be regarded as a binary classification problem, and sampling and forwarding interference has a certain pattern. This means that its data features are relatively obvious, and neural network can be used to extract the features of interference signal. In order to better utilize the continuity information in radar signal, attention mechanism can be used, which uses the context information of input data features for weighted average operation, and dynamically adjusts the attention weights between features through correlation calculation. Since radar intermittent sampling and forwarding interference is sampled from radar signal as source signal, it has a high correlation with radar signal. For the identification of different types of long-delay sampling and forwarding interference signals of radar system, traditional LSTM network will have long-distance dependency problem when processing long data, and attention mechanism can alleviate this problem to a certain extent. In summary, the present invention adopts Transformer network to extract the features of interference signal. Transformer is a deep learning model based on attention mechanism, which is suitable for parallel computing, so that its accuracy and performance are higher than previous neural network.

[0080] Specifically, the Transformer network in step S2 includes: an encoder layer, a decoder layer and a linear output layer connected to each other, wherein the encoder layer is composed of a plurality of encoder blocks, and the decoder layer is composed of a plurality of decoder blocks.

[0081] Specifically, the signal data in step S2 is processed as follows at the encoder layer:

[0082] S2.1.1, the signal data with a length of 1024 is added to the positional encoding after passing through the linear embedding layer. The calculation formula of the positional encoding in the Transformer is as follows:

[0083] (1);

[0084] (2);

[0085] in, and Represents the encoding of the signal at the even position and odd position respectively, and the dimension is the same as Embedding to ensure that it can be added; Indicates the position of a signal data in the entire signal data. express The dimension of Indicates the position of the vector, adds the sin variable to the vector at the even position, adds the cos variable to the vector at the odd position, and finally completes the position encoding operation. The reason why the relative position between signal data can be obtained by using sin encoding and cos encoding is as follows:

[0086] ;

[0087] That is, through By performing a linear transformation, the positional relationship of subsequent data relative to the current data can be obtained.

[0088] S2.1.2, the signal data then enters the encoder block and passes through the multi-head attention mechanism in the encoder block As well as residual connections and layer normalization, the multi-head attention mechanism for:

[0089] (3);

[0090] In the formula, represent The number of columns in the matrix, i.e. the vector dimension, It means to normalize each row of the matrix so that the sum of each row becomes 1. is the query matrix, is the key matrix, is a value matrix.

[0091] S2.1.3, after the feed-forward neural network layer Feed Forward, residual connection and layer normalization Add&Norm, the output of an encoder block is obtained. The output of the previous encoder block is the input of the next encoder block, and the output of the last encoder block is the final output of the encoder layer.

[0092] Specifically, the signal data in step S2 is processed as follows at the decoder layer:

[0093] S2.2.1, same as step S2.1.1, the signal data with a length of 1024 is added to the position encoder after passing through the linear embedding layer.

[0094] S2.2.2, the signal data then passes through the first multi-head attention mechanism in the decoder block. The first multi-head attention mechanism uses a masked operation, and then undergoes residual connection and layer normalization.

[0095] S2.2.3, then passes through the second multi-head attention mechanism. The K and V matrices in the second multi-head attention mechanism are calculated using the output of the encoder layer, while Q is calculated using the output of the previous decoder block.

[0096] S2.2.4, after the forward neural network layer, residual connection and layer normalization, the output of a decoder block is obtained. The output of the previous decoder block is the input of the next decoder block, and the output of the last decoder block is the final output of the decoder layer.

[0097] S2.2.5, the final output of the decoder layer is obtained after the linear layer and the Softmax layer. The output result is in the form of marking the interference position as '1' and the non-interference position as '0'.

[0098] Since the present invention uses the Transformer network for classification operations rather than prediction operations, the input data of the decoder is the same as the input data of the encoder. The role of the Masked Multi-Head Attention mechanism is to make the decoder pay attention to the output of the encoder during the decoding process, ensuring that the data generated by the decoder at the current moment only depends on previous information and avoids using future information. The reason why the second Multi-Head Attention in the decoder block uses the Encoder output to calculate the K and V matrices is that every data in Dncoder can utilize the information of all the data of the Encoder. The subsequent process is similar to that of the Encoder. Finally, the decoder output is processed by Linear and Softmax to obtain the final output. The output result is in the form of marking the interference position as '1' and the non-interference position as '0'.

[0099] Table 1 Transformer network parameters

[0100] .

[0101] In the previous section, we proposed to use the Transformer network to classify radar signals. The proposed method has high detection performance, but there is still one problem that needs to be solved. Due to the sudden change of data features at the junction of interference signals and non-interference signals, the continuity information inside the data cannot be effectively used by the network, which will cause detection errors at the junction of interference signals and non-interference signals.

[0102] In order to deal with such phenomenon, the present invention introduces a CFAR detector to further detect the radar signal. Radar signal detection belongs to a binary detection problem, which means either there is a target or there is no target.

[0103] Specifically, step S3 includes the following steps:

[0104] S3.1, when there is no interference signal input Assumption: When the input signal includes interference signal, radar signal and noise signal, Assume that, expressed as formula (4):

[0105] (4);

[0106] in, is the noise signal, is the radar signal, is an interference signal.

[0107] S3.2, judgment threshold voltage The calculation formula is as follows:

[0108] (5);

[0109] in, is the false alarm rate, is the average power of the non-interference signal, calculated as follows:

[0110] (6);

[0111] In the formula Indicates the signal The amplitude of a signal, is the signal length of the non-interference position in a signal obtained according to the Transformer detection result.

[0112] S3.3, by referring to the detection result of Transformer, we get the target signal without interference and calculate the average power of the target signal according to formula (6). Set , , Three different values ​​are calculated by formula (5) , , Decision threshold voltage under three false alarm rate values , when the voltage value of the detection signal is greater than When it is marked as '1', it means there is interference signal, and when it is less than The mark '0' indicates that there is no interference signal. The detection rule is also the same at the interference boundary, so there will be no detection error due to the interruption of continuity information.

[0113] In the above introduction, the present invention proposes a method of detecting interference using a Transformer network and combining the Transformer detection results to perform CFAR to detect interference. These two methods have their own advantages and disadvantages. The detection results of the Transformer network have a high detection accuracy rate under different interference powers, but detection errors may occur at the edge of the interference; although the CFAR detection results can avoid the phenomenon of detection errors at the edge of the interference and the effect will be better than the Transformer detection results when the interference power is high, the detection results when the interference power is low are much lower than the Transformer detection results. In order to combine the advantages of the two detection results and have good detection results under different interference powers, we need to select different detection results according to the different interference powers to form the final detection results.

[0114] Based on this, the present invention is implemented using a decision tree. A decision tree is a model that displays decision rules and classification results in a tree structure, which can transform seemingly disordered data into a tree model for predicting unknown data. In the prediction process of the decision tree, each path from the root node to the leaf node represents a decision rule, and the class of the leaf node is used as the prediction result. The ultimate goal of decision tree learning is to generate a decision tree with strong generalization ability so that it can predict unknown data.

[0115] Specifically, step S4 includes the following steps:

[0116] S4.1, use the classification and regression tree CART algorithm to create a decision tree. The calculation formula of the Gini coefficient is as follows:

[0117] (7);

[0118] In the formula, represents the number of categories, Represents the probability that a sample belongs to this category, Represents the dataset The probability that two randomly selected samples will have different categories.

[0119] S4.2, for the dataset , according to the characteristics Can you take a value? , divide the sample into two data sets and , satisfying the following relationship:

[0120] (8).

[0121] Specifically, step S4 also includes the following steps:

[0122] S4.3, in the characteristics Under the condition of The Gini coefficient Defined as:

[0123] (9);

[0124] in, , , Represents the data set , and The number of .

[0125] S4.4, select: average power ratio, instantaneous power ratio, Transformer detection results, and CFAR detector false alarm rate , , The three detection results at the time of are used as reference features of the decision tree, and the average power ratio is calculated as follows:

[0126] (10);

[0127] In the formula, It represents the average power of the interference signal in a signal segment. It represents the average power of non-interference signal in a signal. Indicates the average power ratio.

[0128] The instantaneous power ratio is calculated as follows:

[0129] (11);

[0130] In the formula Indicates the number 1 in a signal The amplitude of a signal, It represents the average power of non-interference signal in a signal. Indicates the number of signal points contained in a signal. Indicates The instantaneous power ratio of the signals.

[0131] In order to verify the effectiveness of the proposed method, this embodiment conducted an experiment, which is divided into the following three parts:

[0132] 1) Verification of fusion detection process: Interference detection is performed using simulation data generated by Matlab to verify the feasibility of the method proposed in this invention.

[0133] 2) Statistical performance experiments of different interference types and different JSNRs: In order to verify the detection performance of the proposed method, performance statistical experiments were conducted under different interference types and different JSNRs. The proposed method was then compared with the other two traditional detection methods, proving that the proposed method has a robust detection effect under low JSNRs compared with traditional methods.

[0134] 3) Experiment on improving the robustness of dense interference detection: The present invention conducts statistical performance experiments using the proposed method and the traditional method under different interference duty cycles, and proves that the proposed method is improved over the traditional method under different interference duty cycles.

[0135] First, a simulation experiment is conducted to intuitively verify the feasibility of the method proposed in this invention. In this experiment, the interference signal-to-noise ratio (JSNR) is defined as the ratio of the noise average power to the target average power and noise:

[0136] ;

[0137] In the formula represents the power of the interference signal, represents the power of the radar signal, Indicates the power of the noise signal. In the simulation signal, the radar signal uses a linear frequency modulation (LFM) signal, and the added noise is Gaussian white noise.

[0138] First, the training and verification experiments of the Transformer network are carried out. The simulation signal data is generated by Matlab and is divided into a training data set and a verification data set. The interference type is intermittent sampling direct forwarding interference ISDJ. The training data set consists of 900 sets of data, the interference type is ISDJ, the interference signal-to-noise ratio JSNR range is 5-30dB, 100 sets of data are generated every 3dB interval, each sample length is 1024 sampling points, and its corresponding label data represents the interference position. The verification set has 4500 sets of data, the JSNR range is 6-30dB, 500 sets of data are generated every 3dB interval, the sample length is 1024 sampling points, and its corresponding label data represents the interference position. In order to approach the extreme electromagnetic environment in reality, the sampling period and sampling pulse width in the simulated sampled forwarding interference data are not fixed values, and they will change every time they are sampled. The specific parameters of the signal simulation are shown in Table 2.

[0139] Table 2 Simulation signal parameter settings

[0140] .

[0141] The experiment uses an NVIDIA RTX 2080Ti GPU with 11GB of video memory to train the Transformer neural network. The deep learning framework Pytorch 2.1.1 is used to build the network model. The training batch size is set to 1. The Adam optimizer is used to update the network parameters. The cross entropy loss function is used as the optimization target, which is defined as follows:

[0142] ;

[0143] In the formula, represents the true label of the sample, Represents the label predicted by the model. Set the learning rate , the number of iterations epoch is set to 1000 times, and the change trend of the loss function in the first 100 training processes is as follows Figure 1 As shown. Figure 4As can be seen from the figure, the loss value Loss decreases very quickly in the first 20 trainings and shows a convergence trend. When the number of trainings reaches 40, the Loss value has begun to converge. When the number of trainings reaches 60, the Loss value is close to 0. This shows that the performance of the Transformer network on the training set has improved rapidly.

[0144] The trained network is used to test the validation data, and the detection probability of interference signals is statistically analyzed as the JSNR changes. The detection probability statistics are performed 500 times every 3dB when the JSNR is 6-30dB. The interference detection samples and detection probability statistics under different JSNR are shown in the figure. Figure 2 , Figure 3 , Figure 4 and Figure 5 As shown:

[0145] from Figure 2 , Figure 3 and Figure 4 It can be seen that when JSNR is 6, 9, and 12 dB, the Transformer detection results and the true value of the interference signal position are mostly overlapped, and there are only a few detection errors at the edge of the interference. The reason is that the Transformer recognizes the interference signal based on its density and continuity. At the intersection of interference, the mutation of signal characteristics will cause detection errors. Figure 5 It can be seen that the Transformer detection results have a detection probability of more than 98% under different JSNRs. This is because the Transformer network identifies interference signals by capturing features in the data, and gradually optimizes the parameters in the network through learning each epoch to achieve an optimal state. These features are less affected by the interference power. As the JSNR gradually increases, the detection accuracy is also gradually increasing. When the JSNR reaches 12dB, the interference signal detection accuracy exceeds 99%, but as the JSNR continues to rise, the detection accuracy is still a certain distance from 100%, which is caused by interference edge detection errors. To solve this problem, the present invention uses CFAR detection in the next step.

[0146] Since the Transformer detection results have reliable and robust detection probabilities under different JSNRs, the Transformer detection results are used to estimate the power of the non-interfering signal, which solves the problem of difficulty in selecting the reference unit of the CFAR detector in dense interference. Even if there are a small number of erroneous results, the impact will be reduced when the average power is calculated in formula (6). By setting the false alarm rate to Then, the CFAR detection threshold is calculated according to (5). When the JSNR is 6-30dB, 500 detection probability statistics are performed every 3dB. The detection result samples and detection probability statistics under different JSNR are shown in the figure below. Figure 6 , Figure 7 , Figure 8 and Fig. 9 shown.

[0147] from Figure 6 , Figure 7 , Figure 8 It can be seen that when the JSNR is 6dB, no interference signal is detected, when it is 15dB, some interference signals are detected, and when it is 30dB, the interference signal can be completely detected. Fig. 9 It can be seen that the detection probability of CFAR hardly changes when JSNR is 6-12dB. This is because CFAR is a constant false alarm detection based on power difference. When JSNR is low, the power difference between interference and non-interference signals is not obvious to CFAR, and its calculated detection threshold will be higher than the interference signal, resulting in the interference signal being unable to be detected. As JSNR increases, the power difference between interference and non-interference signals begins to increase, causing the interference signal to gradually exceed the detection threshold of CFAR, which is reflected in the fact that high JSNR (21-30dB) can fully detect the interference signal. When JSNR is 6-12dB, the detection result defaults to 0 when no interference signal is detected, and the true value of interference has only two results, 0 and 1. When comparing, some results will overlap, resulting in a detection probability of 55%-70%.

[0148] Through the previous two steps, it is found that the Transformer network has a good probability of detecting interference signals under different JSNRs, but detection errors may occur at the edge of interference; the CFAR detector that uses the Transformer detection results as a reference to estimate the background power has a 100% detection probability under high JSNR (21-30dB) conditions, but can hardly detect any results under low JSNR (6-12dB) conditions. We want the final detection result to have a higher detection probability of the Transformer under low JSNR and a 100% detection probability of the CFAR detector under high JSNR. To solve this problem, we extracted the interference signal features and combined the two detection results using a decision tree to achieve a robust interference detection probability.

[0149] Before using the decision tree for fusion, you need to first create training data for the decision tree. In the Transformer validation set data used for testing, 100 groups of data are selected under different JSNRs. Perform the following operations on each group of data. First, calculate the average power ratio and instantaneous power ratio corresponding to each group of data, combine the previously obtained Transformer detection results, CFAR detection results under different false alarm rates, and label data, and combine these data. The data is distributed by column for subsequent decision trees to make decisions. Training data The composition is:

[0150] ;

[0151] In the formula is the average power ratio, is the instantaneous power ratio, Transformer detection results, The false alarm rate is CFAR detection results at The false alarm rate is CFAR detection results at The false alarm rate is CFAR detection results at is the true label corresponding to the data. In a set of data, 10 points detected by Transformer as 0 (no interference) and 10 points detected by Transformer as 1 (interference) are randomly selected to form new data, and then the same operation is performed on the remaining 99 sets of data. In this way, training data with a length of 2000 points in each column is obtained under each JSNR. Finally, 9 sets of training data with a length of 18,000 points in each column under different JSNR are obtained. 80% of the total amount of these data are randomly selected, and then one result detected by Transformer as 0 is selected every 10 positions. Finally, the result detected by Transformer as 0 is placed in front, and the result detected by Transformer as 1 is placed in the back, so that it becomes data with a length of 7954 points in each column. Then the detection probability statistics are performed. The generated partial decision tree is as follows Fig.10 As shown, from Fig.10It can be seen that the features selected by the decision tree, such as Transformer results, instantaneous power, etc., as well as the decision strategies under different features. When the Transformer result is <0.5 and the instantaneous power is <8.99124, the decision strategy is 0; when the Transformer result is <0.5 and the instantaneous power is ≥8.99124, the decision strategy is 1; when the Transformer result is ≥0.5 and the instantaneous power is <0.88129, the decision is made according to the average power; when the Transformer result is ≥0.5 and the instantaneous power is ≥0.88129, the decision is made according to the instantaneous power. When the JSNR is 6-30dB, 500 detection probability statistics are performed every 3dB. Sample detection results under different JSNRs are shown below. Fig.11 , Fig.12 and Fig.13 As shown, from Fig.11 It can be seen that when JSNR is 6dB, the detection probability is relatively Figure 2 Transformer detection probability decreases slightly, but is comparable to that of the JSNR at 15dB and 30dB. Figure 3 , Figure 4 Compared with the detection error at the interference edge, the detection probability of the decision tree is eliminated. The detection probability of the decision tree is statistically analyzed and compared with the Transformer result and the CFAR result to intuitively prove that the fusion result is improved. The results are as follows: Fig.14 As shown in the figure, when JSNR is 6dB, the fusion detection probability of the decision tree is slightly lower than that of the Transformer. This is because the power difference between the interference signal and the non-interference signal is not obvious at 6dB, resulting in very poor CFAR detection results and almost no interference detection. There is also very little effective decision information when making decisions in the decision tree, which will lead to misjudgment when the decision tree makes decisions. From 9dB, the fusion detection probability of the decision tree has begun to be higher than that of the Transformer and CFAR. This is because the increase in interference power has a more obvious difference in the power of the interference signal and the non-interference signal. Even when the CFAR detection probability is low at JSNR of 9-15dB, the combination of the Transformer detection results and the power characteristics in the signal enables the decision tree to make a correct decision. At this time, the detection probability has exceeded 99%. When JSNR is 21-30dB, the CFAR detection probability has reached 100%, exceeding the Transformer detection probability. At this time, the detection probability obtained after fusion has almost reached 100%.

[0152] The above results show that the method proposed in this invention can effectively detect interference signals. The final fusion results show that the method proposed by us has robust detection performance for interference signal detection, which proves the effectiveness of the detection method proposed by this invention.

[0153] Next, statistical performance experiments with different interference types and different JSNRs will be conducted to verify the robustness of the method proposed in the present invention compared with the traditional method under low JSNR.

[0154] In this part, the training data set and verification data set used by the present invention are composed of three interference types: ISDJ, ISRJ, and ISCJ. The training data set consists of 2700 groups of data, 900 groups of data for each interference type, and 100 groups of data are generated every 3dB for each interference type in JSNR5-30dB. Each sample length is 1024 sampling points, and its corresponding label data represents the interference position. The verification set has 4500 groups of data for each interference type, JSNR range is 6-30dB, 500 groups of data are generated every 3dB, the sample length is 1024 sampling points, and its corresponding label data represents the interference position. The signal simulation parameters are shown in Table 2, rows 2-4. The sampling pulse width and sampling period value ranges of the three interferences are shown in Table 3:

[0155] Table 3 Pulse width and sampling period value ranges for different interference types

[0156] .

[0157] Next, the detection probability experiment was carried out using the method proposed in the present invention and two traditional methods under different interference types and different JSNR conditions. Statistical experiments were performed 500 times every 3dB in the JSNR range of 6-30dB. Among them, when the interference type is ISDJ and the JSNR is 6dB, the method proposed in the present invention has obvious detection errors but a higher overall detection probability, while the other two traditional detection methods did not detect any interference signals. When the JSNR is 9dB, the method proposed in the present invention has only very few detection errors, while the two traditional detection methods still cannot detect any interference. When the JSNR is 12dB, the method proposed in the present invention can detect all interference signals, and the Adaptive Threshold CFAR method can also detect some interference signals intermittently, but there are still some interference signals that are not detected, and the Bi-Directional CFAR method still does not detect any interference signals. In order to more intuitively show the difference in detection probability between the method proposed in the present invention and the traditional method under low JSNR, the method proposed in the present invention can achieve a detection probability of more than 96.5% when the JSNR is 6-12dB under three different types of interference, while the detection results of the traditional method either cannot detect interference or can only detect part of the interference under the same JSNR conditions. The detection probability of the method proposed in the present invention is 10%-40% higher than that of the traditional method. This is because the traditional CFAR detection method always focuses on the power difference between a single detection unit and the background power, which makes it invalid under low JSNR conditions, and due to the existence of the protection unit, the power distribution near the detection unit cannot be considered. The method proposed in the present invention improves the detection robustness under low JSNR conditions by fusing the feature detection framework. The use of the continuous distribution characteristics of the interference signal solves the problem of difficulty in interference detection under low JSNR conditions. Compared with the traditional method, the process of Transformer preliminary detection and then CFAR detection avoids the problem of difficulty in selecting reference units. In short, compared with the traditional method, the detection probability of this method under low JSNR is robust, and it has a robust detection effect under different types of interference.

[0158] Finally, statistical experiments are performed using the method proposed in the present invention and the traditional method under different interference duty cycles to prove that the method proposed in the present invention has a robust performance improvement compared with the traditional method under dense interference.

[0159] First, define the concept of interference duty cycle, which is calculated based on the ratio of the interference signal length to the total signal length. The calculation formula is as follows:

[0160]

[0161] in, is the interference signal length, is the total length of the signal, is the calculated interference duty cycle.

[0162] When the interference duty cycle is 25%, 45% and 60%, the interference type is ISDJ, and the JSNR is 9dB, the method proposed in the present invention can detect the interference well, and only individual positions have detection errors. The traditional method can only detect partial interference or cannot detect interference at all. When the interference becomes more intensive, the interference that the traditional method can detect is further reduced. When the density of the interference increases further, the traditional method cannot detect interference at all. This is because as the interference becomes more intensive, it becomes more difficult for the traditional method to select a pure and interference-free reference unit, affecting the detection effect. Under these three interference duty cycles, the method proposed in the present invention and the traditional method were used to perform 500 detection probability statistical experiments at each JSNR of 6, 9, 21, and 30dB for the interference type. The results are shown in Table 4.

[0163] Table 4 Detection results under different interference duty cycles

[0164] .

[0165] It can be seen from Table 4 that with the increase of interference duty cycle, the detection probability of traditional detection methods under different JSNRs is affected. When JSNR is 30dB under low interference duty cycle, both traditional detection methods can achieve 100% detection probability. When JSNR is 30dB under medium interference duty cycle, the performance of both traditional methods is affected and the detection probability decreases by 1%-2%. At high interference duty cycle, the detection performance of traditional methods is greatly affected when JSNR is 30dB, and the detection probability decreases significantly. The detection probability of the method proposed in the present invention has consistent detection results under different interference duty cycles. This is because when the interference distribution is dense, the traditional CFAR detection method mixes high-power interference signals into the reference unit used for background power estimation, resulting in an increase in the detection threshold and missed detection results. The method proposed in the present invention has a hierarchical detection framework, which is divided into three layers. First, Transformer is used for rough detection to assist in selecting reference units; secondly, the rough classification results of interference and non-interference signals can be used to estimate the power of non-interference signals to solve the difficulty of selecting reference units under strong interference. The CFAR detection result is obtained by estimating the power of the non-interference signal; finally, the interference signal features are extracted, and a probability statistics and multi-feature fusion detection model is proposed to achieve robust interference detection. The hierarchical detection framework solves the difficulties of dense interference detection by decomposing a large task into several sub-detection problems. Compared with traditional methods, the detection results of Transformer avoid the difficulties of reference unit selection. Therefore, these results show that the method proposed in the present invention has better detection results and robust performance improvement compared with traditional methods under dense interference.

[0166] In summary, the traditional interference detection method has a good interference detection effect when the JSNR is high and the interference is not dense. However, when the JSNR is low and the interference is dense, the detection effect is not ideal. The present invention proposes a radar interference detection method based on CFAR and decision tree. The proposed interference detection method has the advantage of high detection accuracy under different interference intensities and different JSNR conditions. Further work can suppress the identified interference signals and enhance the anti-interference capability of the radar.

[0167] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A radar intermittent sampling and forwarding interference detection method based on a decision tree, characterized in that: The specific steps include: S1, standardize the radar signal, including: modulus normalization and length clipping; S2, using the Transformer network to detect interference signals and obtain the detection results of the Transformer network; S3, using the detection results of Transformer as a reference, distinguishes interference signals from non-interference signals, calculates the detection threshold, and uses the CFAR detector to further detect interference signals; S4, using a decision tree to integrate the detection results obtained in steps S2 and S3, the instantaneous power and the average power of the signal, and adopting different result selection strategies under different interference powers to obtain the final detection result; Step S3 specifically includes the following steps: S3.1, when there is no interference signal input Assumption: When the input signal includes interference signal, radar signal and noise signal, Assume that, expressed as formula (4): (4); in, is the noise signal, is the radar signal, is the interference signal; S3.2, judgment threshold voltage The calculation formula is as follows: (5); in, is the false alarm rate, is the average power of the non-interference signal, calculated as follows: (6); In the formula Indicates the signal The amplitude of a signal, is the signal length of the non-interference position in a signal obtained according to the Transformer detection result; S3.3, calculated by formula (5) , , Decision threshold voltage under three false alarm rate values , when the voltage value of the detection signal is greater than When it is marked as '1', it means there is interference signal, and when it is less than When marked as '0', it means there is no interference signal; Step S4 specifically includes the following steps: S4.1, use the classification and regression tree CART algorithm to create a decision tree. The calculation formula of the Gini coefficient is as follows: (7); In the formula, represents the number of categories, Represents the probability that a sample belongs to this category, Represents the dataset The probability that two randomly selected samples have different categories; S4.2, for the dataset , according to the characteristics Can you take a value? , divide the sample into two data sets and , satisfying the following relationship: (8); Step S4 also includes the following steps: S4.3, in the characteristics Under the condition of The Gini coefficient Defined as: (9); in, , , Represents the data set , and The number of S4.4, select: average power ratio, instantaneous power ratio, Transformer detection results, and CFAR detector with false alarm rate , , The three detection results at the time of are used as reference features of the decision tree, and the average power ratio is calculated as follows: (10); In the formula, It represents the average power of the interference signal in a signal segment. It represents the average power of non-interference signal in a signal. represents the average power ratio; The instantaneous power ratio is calculated as follows: (11); In the formula Indicates the number 1 in a signal The amplitude of a signal, It represents the average power of non-interference signal in a signal. Indicates the number of signal points contained in a signal. Indicates The instantaneous power ratio of the signals.

2. According to a decision tree-based radar intermittent sampling and forwarding interference detection method according to claim 1, it is characterized in that: The Transformer network in step S2 includes: an encoder layer, a decoder layer and a linear output layer connected to each other, wherein the encoder layer is composed of a plurality of encoder blocks, and the decoder layer is composed of a plurality of decoder blocks.

3. According to a decision tree-based radar intermittent sampling and forwarding interference detection method according to claim 2, it is characterized in that: In step S2, the signal data is processed at the encoder layer as follows: S2.1.1, the signal data with a length of 1024 is added to the position encoder after passing through the linear embedding layer, where the calculation formula of the position encoder in the Transformer is as follows: (1); (2); in, and Represent the codes of the signal at even and odd positions respectively. Indicates the position of a signal data in the entire signal data. express The dimension of Represents the position of a vector; S2.1.2, the signal data then enters the encoder block and passes through the multi-head attention mechanism in the encoder block As well as residual connections and layer normalization, the multi-head attention mechanism for: (3); In the formula, represent The number of columns in the matrix, i.e. the vector dimension, Indicates that each row of the matrix is ​​normalized, where is the query matrix, is the key matrix, is the value matrix; S2.1.3, after the forward neural network layer, residual connection and layer normalization, the output of an encoder block is obtained. The output of the previous encoder block is the input of the next encoder block, and the output of the last encoder block is the final output of the encoder layer.

4. According to the decision tree-based radar intermittent sampling and forwarding interference detection method of claim 3, it is characterized in that: In step S2, the signal data is processed in the decoder layer as follows: S2.2.1, same as step S2.1.1, the signal data with a length of 1024 is added to the position encoder after passing through the linear embedding layer; S2.2.2, the signal data then passes through the first multi-head attention mechanism in the decoder block, which uses a mask operation, and then undergoes residual connection and layer normalization; S2.2.3, then passes through the second multi-head attention mechanism, the K and V matrices in the second multi-head attention mechanism are calculated using the output of the encoder layer, and Q is calculated using the output of the previous decoder block; S2.2.4, after the forward neural network layer, residual connection and layer normalization, the output of a decoder block is obtained. The output of the previous decoder block is the input of the next decoder block, and the output of the last decoder block is the final output of the decoder layer; S2.2.5, the final output of the decoder layer is obtained after the linear layer and the Softmax layer. The output result is in the form of marking the interference position as '1' and the non-interference position as '0'.