PDW Homology Matching Method Based on Self-Attention Mechanism

Abstract

The present invention discloses a PDW homologous matching method based on a self-attention mechanism, which includes the following steps; Step 1: Construct a data set, and divide the data set into the following three parts according to the division of the matching scenario: synchronous cross-system data, asynchronous cross-system data, and asynchronous same-system data; Step 2: Data preprocessing, which is the preprocessing of the data in Step 1 before entering the algorithm; Step 3: Type judgment, respectively perform type judgment on the matching parameters of the data, namely carrier frequency RF, pulse width PW, and pulse PRI; The data types include non-agile, periodic agile, and non-periodic agile; Step 4: Construct a self-attention mechanism network to judge whether two groups of RF or PW or PRI that are both periodic agile or non-periodic agile are homologous. The present invention can perform homologous matching on asynchronous radar PDW strings, utilize the unexploited positioning parameters within a single subsystem to improve the positioning success rate of the integrated system.

Description

Technical Field

[0001] The present invention belongs to the technical field of PDW homologous matching, and particularly relates to a PDW homologous matching method based on a self-attention mechanism. Background Art

[0002] Conventional data association and homologous matching based on PDW (pulse description word) are commonly found in the pre-sorting processing of distributed radar time difference sorting and the radar batch combination algorithm.

[0003] In a multi-station electronic reconnaissance system, the main task of the pre-sorting processing is to cluster the radiation source signals with different characteristics from multiple dimensions such as the frequency domain, spatial domain, time domain, and energy domain according to the characteristics of the radar signals and the prior knowledge of the known radars, and assign the PDW (pulse description word) of the radar signals to different groups, reducing the processing burden of the main sorting algorithm and laying a foundation for subsequent sorting. Traditional pre-sorting methods include the correlation comparator method, clustering algorithms such as the KNN algorithm, the K-means clustering algorithm, etc. The pre-sorting processing of radar multi-station time difference sorting is to match the same pulse received by two observation stations to form pulse pair data. The constraint conditions for pulse pairing include the time difference window constraint condition and the PDW (pulse description word) parameter constraint condition. The clustering algorithm mainly calculates the similarity of radar pulse signals according to the difference of radar PDW (pulse description word), and classifies the samples through certain judgment criteria. The clustering algorithm generally judges the similarity by calculating the distance between samples, and common distance expressions include cosine distance, correlation distance, Hamming distance, Mahalanobis distance, etc. Current multi-parameter clustering algorithms include K-means clustering, hierarchical clustering, density clustering, grid clustering, fuzzy clustering, clustering algorithms based on swarm intelligence optimization, and so on.

[0004] Radar batch combination is mainly applied to the situation where signals belonging to a radar radiation source are sorted into several due to complex electromagnetic environment and noise interference. At this time, it is necessary to perform batch combination processing on the radar sorting results. The commonly used traditional algorithm in radar batch combination is pattern recognition based on the membership function. The confidence of the radar pulse to be matched and the known radar working mode is calculated by establishing a membership function.

[0005] Most algorithms related to radar PDW (pulse description word) matching require two groups of PDW (pulse description word) data strings within the same time difference window.

[0006] However, in the scenario of the research problem in the integrated positioning and monitoring system, with the enemy's radar radiation source as the processing object, radar pulse sorting is completed separately within two subsystems. However, at this time, due to reasons such as terrain occlusion, limited monitoring distance, and high target mobility within a single subsystem, there is non-common visibility of the target, resulting in the positioning parameters measured within a single subsystem at this moment being insufficient to support time difference positioning. When both subsystems are unable to perform positioning due to non-common visibility, the radar PDW (pulse description word) data at similar times sorted by the two subsystems are subjected to homologous matching, and based on this, the one-way time difference of arrival (TDOA) that cannot be positioned measured by the two subsystems is fused and utilized for positioning. At this time, the PDW (pulse description word) strings used for matching come from different subsystems. Due to the different distances between the target and the stations of different subsystems, it is very difficult to ensure that the PDW (pulse description word) strings used for matching belong to the same time difference window, that is, the two groups of PDW (pulse description word) strings are asynchronous.

[0007] Traditional algorithms are difficult to meet the requirements of asynchronous data matching. Summary of the Invention

[0008] In order to overcome the deficiencies of the above-mentioned existing technologies, the purpose of the present invention is to provide a PDW homologous matching method based on the self-attention mechanism, which can perform homologous matching on asynchronous radar PDW (pulse description word) strings, utilize the positioning parameters not utilized within a single subsystem, and thereby improve the positioning success rate of the integrated system.

[0009] In order to achieve the above purpose, the technical solution adopted by the present invention is:

[0010] A PDW homologous matching method based on the self-attention mechanism includes the following steps;

[0011] Step 1: Construct a data set, and divide the data set into the following three parts according to the division of the matching scenario: synchronous cross-system data, asynchronous cross-system data, and asynchronous same-system data;

[0012] Step 2: Data preprocessing, preprocess the data in Step 1 before entering the algorithm, and use it to extract and calculate the matching parameters carrier frequency, pulse width, and pulse repetition interval for use by the matching algorithm;

[0013] Step 3: Type judgment, respectively perform type judgment on the matching parameters carrier frequency RF, pulse width PW, and pulse PRI of the data; the data types include non-agile, periodic agile, and non-periodic agile;

[0014] If the data types of two groups of RF, or two groups of PW, or two groups of PRI are different, it is directly determined as non - homologous. If two groups of RF, or two groups of PW, or two groups of PRI are not agile, the Euclidean distance is directly calculated to determine whether they are homologous. If two groups of RF, or two groups of PW, or two groups of PRI are all periodically agile or non - periodically agile, they are sent to the self - attention mechanism network constructed in step 4 to determine whether they are homologous;

[0015] Step 4: Construct a self - attention mechanism network to determine whether two groups of RF, or PW, or PRI with all being periodically agile or non - periodically agile are homologous.

[0016] In the scenario of synchronously cross - system data in step 1, the target signal is intercepted by two subsystems at positions at almost the same moment. At this time, the pulses from the two subsystems will have partial or complete overlap. There is a pulse received from site B within the pulse time - difference window received at site A i ; Synchronously cross - system data consists of homologous agile synchronously cross - system data, homologous non - agile synchronously cross - system data, non - homologous agile synchronously cross - system data, and non - homologous non - agile synchronously cross - system data; j In the scenario of constructing asynchronously cross - system data, the target signal is intercepted by two subsystems at positions at close but different moments successively. At this time, the pulses from the two subsystems are difficult to align. There is basically no pulse received from site B within the pulse time - difference window received at site A

[0017] ; Asynchronously cross - system data consists of homologous agile asynchronously cross - system data, homologous non - agile asynchronously cross - system data, non - homologous agile asynchronously cross - system data, and non - homologous non - agile asynchronously cross - system data; i In the scenario of constructing asynchronously same - system data, the data set comes from PDW (Pulse Description Word) data received by the same subsystem at different moments. After positioning is completed, the PDW (Pulse Description Word) parameter information obtained this time is matched with the PDW (Pulse Description Word) parameter information of known target batch numbers within the same subsystem to complete the function of target batch numbering; Asynchronously same - system data consists of homologous agile asynchronously same - system data, homologous non - agile asynchronously same - system data, non - homologous agile asynchronously same - system data, and non - homologous non - agile asynchronously same - system data j ;

[0018] Specifically, step 2 is as follows:

[0019] To extract the inter - pulse modulation information of the monitored target, the pulse repetition interval PRI is calculated from the TOA measured by the system monitoring, and the calculation method is as follows:

[0020] PRI(k)=TOA(k)-TOA(k - 1), k ∈[2,N]

[0021] ​

[0022] Where N is the number of pulses, PRI is the pulse repetition interval, and TOA is the time of arrival;

[0023] After calculating the PRI for the two sets of radar pulse description word data obtained from system monitoring measurements respectively, extract the carrier frequency, pulse width, and pulse repetition interval of the two sets of PDW (pulse description word) parameters, and combine them into three sets of data with a dimension of N×2. N is the number of pulses. The three parameter matrices are in the form as shown in the following formula:

[0024]

[0025] Where, with RF 1n Note that the subscript 1 indicates that the parameter comes from the first group, and the subscript n is the carrier frequency value of the nth pulse of this group of parameters.

[0026] In step 3, if neither the two sets of RF, nor the two sets of PW, nor the two sets of PRI are agile, directly compare them to calculate the Euclidean distance to determine whether they are the same; if the two sets of parameter types are different, directly determine them as different; if both sets of parameters are determined to be periodically agile, send them to the network for periodically agile data for matching to determine whether they are the same; if both sets of parameters are determined to be non-periodically agile, send them to the network for non-periodically agile data for matching to determine whether they are the same. The three main matching parameters RF, PW, and PRI all go through the above process. If all three parameters can be matched as the same, it is determined to be homologous, otherwise it is determined to be non-homologous.

[0027] In step 3, perform type judgments on the matching parameters of the data, namely the carrier frequency (RF1, RF2), pulse width (PW1, PW2), and pulse (PRI1, PRI2). Here, the subscript 1 and subscript 2 represent that the parameters come from two different stations.

[0028] The agility and periodicity judgment in step 3 is specifically as follows:

[0029] (3.1) Agility judgment

[0030] Before judging whether the parameter is agile, first perform Min-max normalization on the parameter (the carrier frequency, pulse width, and pulse repetition interval within the PDW parameter), and map the value to between [0, 1]. The formula is as shown below:

[0031]

[0032] Where x is the parameter sequence, min(x) and max(x) are the minimum and maximum values of the parameter sequence x respectively, and x * is the value of the parameter sequence x after normalization;

[0033] After normalization, calculate the standard deviation of the parameter, and use the standard deviation to measure whether the data parameter is agile;

[0034] (3.2) Periodic judgment;

[0035] Judgment is made through the autocorrelation result, and the autocorrelation formula is as follows:

[0036]

[0037] where x(n) is the parameter sequence, x(n - m) is the value of x(n) after being shifted by m, N is the number of parameters in the parameter sequence, and φ x (m) is the autocorrelation result;

[0038] Since the time-domain autocorrelation is equivalent to the frequency-domain conjugate multiplication, and under the same amount of data, the operation speed of the frequency-domain conjugate multiplication is much higher than that of the time-domain autocorrelation. Therefore, the autocorrelation is calculated by the frequency-domain conjugate multiplication of the time domain, and the formula is as follows:

[0039] φ x (m) = ifft(fft(x(n)) * conj(fft(x(n))))

[0040] where x(n) is the parameter sequence, fft is the fast Fourier transform, ifft is the inverse fast Fourier transform, conj is the complex conjugate operation, and φ x (m) is the autocorrelation result;

[0041] The autocorrelation result of the periodic sequence will have multiple highly approximate peaks. Based on this feature, the parameters with agility are divided into periodic agility and non-periodic agility. Among them, the periodic agility includes PRI slide, PRI stagger, and frequency diversity, and the non-periodic agility includes PRI jitter.

[0042] The process of feeding the periodic agility data and non-periodic agility data into the network to obtain the homologous result in step 4 is as follows:

[0043] First, the carrier frequency, pulse width, or pulse repetition interval of two groups of PDW in the parameter data is normalized by Z_score, and the formula is as follows:

[0044]

[0045] where μ is the mean of all sample data, σ is the standard deviation of all sample data, and x is the carrier frequency (RF1, RF2) or pulse width (PW1, PW2) or pulse repetition interval (PRI1, PRI2);

[0046] Then, the standardized data is fed into the pre-trained self-attention mechanism network, which mainly consists of the following two parts: a similarity calculation module and a classification module. The data first passes through the similarity calculation module composed of three layers of self-attention mechanisms. The self-attention mechanism can measure the similarity between the input sample itself and itself through the operations of the input Q, K, and V matrices. Taking the input data as the carrier frequency as an example, the similarity calculation module is calculating the similarity between (RF1, RF2) at this time. After obtaining the similarity matrix, it is sent to the classification module, which consists of a fully connected layer and an activation function. Judging whether they are from the same source can essentially be regarded as a classification problem, and the classification is based on the similarity matrix to obtain the homologous matching result of the parameters.

[0047] Advantages of the present invention:

[0048] The present invention takes radar emitters as the processing target. Considering the situation where multiple individual monitoring subsystems cannot locate, for the PDW (pulse description word) data strings received by different monitoring subsystems in the integrated system at similar times, homologous matching is performed. After successful matching, the unutilized positioning parameters in a single subsystem can be utilized to improve the positioning success rate of the integrated system. The research on same-target matching based on PDW (pulse description word) data strings proposed by the present invention also provides a new research method for data fusion of different monitoring platforms. There is little PDW (pulse description word) homologous matching in similar scenarios in existing research.

[0049] The present invention introduces the self-attention mechanism. The self-attention mechanism can obtain the effect of calculating the similarity of the input through the matrix multiplication operations of the Q, K, and V matrices. For example, the inner product of vectors represents the included angle between two vectors and represents the projection of one vector on another vector. A large projection value means a high correlation between the two vectors. The operations of the Q, K, and V matrices in the self-attention mechanism are just like the inner product of two vectors, which can reflect the similarity of the Q, K, and V matrices. Moreover, the Q, K, and V matrices are all calculated from the input matrix, so the self-attention mechanism can measure the similarity of the input parameter matrix. Description of the Drawings

[0050] Figure 1 is the flow chart of the PDW homologous matching algorithm based on the self-attention mechanism of the present invention.

[0051] Figure 2 is the scenario for constructing a synchronous cross-system data set.

[0052] Figure 3 is the pulse schematic diagram of the synchronous cross-system data set.

[0053] Figure 4 is the scenario for constructing an asynchronous cross-system data set.

[0054] Figure 5For the scenario of asynchronous cross - system dataset construction.

[0055] Figure 6 For the scenario of asynchronous intra - system dataset construction.

[0056] Figure 7 For the scenario of asynchronous intra - system dataset construction.

[0057] Figure 8 For the data pre - processing process of the PDW homologous matching network model based on the self - attention mechanism.

[0058] Figure 9 For the comparison chart of the autocorrelation results of periodic and aperiodic sequences.

[0059] Figure 10 For the network model structure diagram.

[0060] Figure 11 For the comparison chart of the matching performance of different algorithms under different parameter measurement noises. Specific implementation manners

[0061] The present invention will be further described in detail below with reference to the accompanying drawings.

[0062] As Figure 1 shown: In view of the fact that the traditional method and the method described in the previous section are not robust enough when facing asynchronous cross - system data and asynchronous intra - system data, this application intends to introduce a self - attention mechanism network to improve the matching success rate of asynchronous data. Datasets of asynchronous cross - system data, asynchronous intra - system data and synchronous cross - system datasets are constructed as the training sets of the self - attention mechanism network. Due to the operations of the internal Q, K, V matrices of the self - attention mechanism, it itself has the function of calculating the similarity of the network input, so it is applicable to the co - visibility association problem of PDW (pulse description word) parameters. Compared with the CNN network, the self - attention mechanism is more suitable for processing time - series data and can better focus on the rapid - change characteristics of the data. Compared with the RNN, the self - attention mechanism pays more attention to the correlation of the input samples themselves. Test verification shows that the self - attention mechanism can maintain a high matching accuracy when facing asynchronous cross - system data and asynchronous intra - system data. The self - attention mechanism network model proposed in this application can take into account the matching of asynchronous cross - system data and asynchronous intra - system data and still has a high matching accuracy when facing asynchronous and rapidly - changing data.

[0063] However, sending all types of datasets to the network may result in a wide network coverage but low accuracy, and the parameter information unrelated to matching, such as amplitude, included therein will also interfere with the network's accuracy. Therefore, the main parameters required for matching, namely RF, PW, and PRI, are extracted, and the RF, PW, and PRI of the two subsystems are respectively sent to the network for matching. The network only needs to determine whether the two sets of parameters sent are from the same signal. So the data is classified into three categories: non-agile, periodically agile such as PRI sliding, PRI staggering, frequency diversity, etc., and non-periodically agile such as PRI jitter, etc., and the network is trained separately for periodically agile and non-periodically agile data to improve the matching accuracy rate.

[0064] (1) Construct the dataset

[0065] The research scenario of this application can be divided into the following three parts: synchronous cross-system data, asynchronous cross-system data, and asynchronous same-system data. The construction of the dataset will also be classified according to the research scenario.

[0066] (1.1) Synchronous cross-system data

[0067] The scenario of synchronous cross-system data is as Figure 2 , Figure 3 shown. The target signal is intercepted by the two subsystems at positions almost at the same moment. At this time, some or all of the pulses from the two subsystems will overlap. There are pulses received by site B within the pulse time difference window received by site A i . j The pulses received by site B.

[0068] (1.2) Asynchronous cross-system data

[0069] The construction scenario of asynchronous cross-system data is as Figure 4 , Figure 5 shown. The target signal is intercepted by the two subsystems successively at positions close but different moments. At this time, it is difficult to align the pulses from the two subsystems. There are basically no pulses received by site B within the pulse time difference window received by site A i . j The pulses received by site B.

[0070] (1.3) Asynchronous same-system data

[0071] The construction scenario of asynchronous same-system data is as Figure 6 , Figure 7As shown, the data set comes from PDW (Pulse Description Word) data received by the same subsystem at different times. When there are multiple targets in the monitoring area, in order to achieve the continuity of trajectory fitting, it is necessary to determine which target in the monitoring area the obtained position information belongs to after the asynchronous cross-system data is matched and the positioning algorithm completes the positioning. Therefore, after the positioning is completed, it is necessary to match the PDW (Pulse Description Word) parameter information obtained this time with the PDW (Pulse Description Word) parameter information of the known target batch numbers within the same subsystem, so as to complete the function of target batch numbering.

[0072] (2) Data preprocessing

[0073] To extract the inter-pulse modulation information of the monitoring target, the pulse repetition interval PRI is calculated using the TOA obtained by the system monitoring measurement. The calculation method is as follows:

[0074] PRI(k) = TOA(k) - TOA(k - 1), k ∈ [2, N]

[0075] where N is the number of pulses.

[0076] As Figure 8 shown: After calculating the PRI for two groups of radar pulse description word data obtained by the system monitoring measurement respectively, the carrier frequency, pulse width and pulse repetition interval of the two groups of PDW (Pulse Description Word) parameters are extracted and combined into three groups of data with a dimension of N×2. N is the number of pulses. The matrix form of the three groups of parameters is shown as follows:

[0077]

[0078] where, taking RF ij as an example, the subscript i indicates that the parameter comes from the i-th group, and the subscript j is the carrier frequency value of the j-th pulse of this group of parameters.

[0079] (3) Type judgment

[0080] (3.1) Agility judgment

[0081] Before judging whether the parameter is agile, first perform Min-max normalization on the parameter. Min-max normalization is a linear transformation of the data, mapping the value to between [0, 1]. The formula is as follows:

[0082]

[0083] After normalization, calculate the standard deviation of the parameter. The standard deviation refers to the average of the squared differences between each sample value and the average of all sample values, which can measure the degree of deviation of a set of data from its expectation and can reflect the dispersion degree of a set of data. Therefore, this application selects the variance to measure whether the data parameter is agile.

[0084] Using Min-max normalization, first, it can unify the threshold range. Second, Min-max normalization can better maintain the data distribution of the original data and is very sensitive to the existence of outliers. Compared with Z-Score standardization, if Z-Score standardization is used, the data distribution will be changed, and the variance calculated at this time is difficult to reflect the degree of data dispersion.

[0085] (3.2) Periodicity judgment

[0086] Judgment is made through the autocorrelation result, and the autocorrelation formula is as follows:

[0087]

[0088] The result of autocorrelation can show the similarity between the sequence x(n) and the value x(n - m) after it is shifted by m.

[0089] In practical applications, if direct time-domain autocorrelation is performed, the operation speed is slow, which will increase the running time of the algorithm. Since time-domain autocorrelation is equivalent to frequency-domain conjugate multiplication, at the same data volume, the operation speed of frequency-domain conjugate multiplication is much higher than that of time-domain autocorrelation. Therefore, time-domain and frequency-domain conjugate multiplication is used to calculate autocorrelation, and the formula is as follows:

[0090] φ x (m) = ifft(fft(x(n)) * conj(fft(x(n))))

[0091] Such as Figure 9 As shown: The autocorrelation results of periodic sequences will have multiple highly approximate peaks. Through this feature, the parameters with agility can be divided into periodic agility and non-periodic agility. Among them, periodic agility includes PRI slide, PRI stagger, frequency diversity, etc., and non-periodic agility includes PRI jitter, etc.

[0092] (4) Construct a self-attention mechanism network

[0093] (4.1) Data standardization

[0094] When the self-attention mechanism network is training to achieve better training results, it is necessary to standardize the input data. In the field of machine learning, different evaluation metrics (that is, different features in the feature vector are the different evaluation metrics mentioned above) often have different dimensions and dimension units. Such a situation will affect the results of data analysis. In order to eliminate the influence of dimensions between metrics, it is necessary to perform data standardization processing to solve the comparability between data metrics. After the original data is processed by data standardization, each metric is at the same order of magnitude, which is suitable for comprehensive comparative evaluation. The standardization algorithm uses the Z-score standardization method. After data processing, it conforms to the standard normal distribution, that is, the mean is 0 and the standard deviation is 1, and its transformation function is:

[0095]

[0096] Among them, μ is the mean of all sample data, and σ is the standard deviation of all sample data. When it is necessary to use distance to measure similarity in classification and clustering algorithms, Z-score standardization performs better.

[0097] (4.2) Loss function

[0098] The cross-entropy loss function, also known as the Softmax loss function, is the most commonly used classification objective function in current neural networks. Its form is:

[0099]

[0100] Among them, M is the number of categories, y ic is the sign function (0 or 1), which takes 1 if the true category of sample i is equal to c, and 0 otherwise, and p ic is the predicted probability that the observed sample i belongs to category c.

[0101] (4.3) Network model optimization

[0102] ① Vanishing gradient and exploding gradient problems

[0103] Vanishing gradient means that the derivative value of the activation function in the convolutional neural network is too small. When the derivative value approaches zero, the weight parameters in the network cannot be updated anymore. Exploding gradient is because as the number of layers increases, the gradient update information increases exponentially, the weight parameter values are too large, and the update speed is too fast. To address such problems, an LN layer is added. The main role of normalization is to normalize each layer of features before they are input into the activation function, converting them into data with a mean of 1 and a variance of 0, so as to avoid the data falling into the saturation region of the activation function and reduce the problem of vanishing gradient.

[0104] LN normalizes all features of each sample for each sample. Compared with BN, BN flattens the magnitude relationship between different features while retaining the magnitude relationship between different samples. In this way, if the specific task depends on the relationship between different samples, BN is more effective, especially in the CV field. For example, when classifying different image samples, the magnitude relationship between different samples is retained. LN flattens the magnitude relationship between different samples while retaining the magnitude relationship between different features. Therefore, LN is more suitable for tasks in the NLP field, where the features of a sample are actually different word embeddings, and the LN can retain this temporal relationship between features.

[0105] The issues discussed in this application do not focus on the differences and connections between different samples, but rather on the relationships between different features within the same sample. Therefore, using LN is more suitable for the issues discussed in this application.

[0106] ② Overfitting phenomenon

[0107] Overfitting is a very common problem in the deep learning process. The essence of overfitting is that the model learns too precisely. Due to the errors, noises, and other interferences in the training data, and the data fitting ability of the neural network model is different from that of traditional functions. During the training process, the model will learn and fit the errors in the data, resulting in good performance of the model during the training process but poor performance during the testing process, making the model only effective for the training dataset, that is, the generalization ability of the model is poor.

[0108] For the solutions to the overfitting phenomenon, it is usually possible to handle it from three aspects: obtaining more data; adjusting the number of model layers and parameters to make the model more suitable; combining multiple models.

[0109] Regarding the possible overfitting problems that may occur during the research process, this application first attempts to use data augmentation methods such as data scaling to increase the training data, and secondly adjusts the model complexity to avoid overfitting caused by the model being too complex.

[0110] ③ Selection of training method and optimizer

[0111] The training method uses mini-batch gradient descent in gradient descent method. Select a bath size data volume between 1 and the maximum training data volume for training. It is a compromise method between batch gradient descent and stochastic gradient descent. Using a batch of data each time to optimize the neural network parameters can greatly reduce the time required for convergence and achieve parallelization.

[0112] The optimizer selects Adam. Adam is a method that adapts different learning rates for different parameters. Adam incorporates the idea of moment estimation and adjusts the hyperparameters in real time by calculating and correcting the first moment and second moment of each round of gradient. It is simple to implement, computationally efficient, and has low memory requirements. After tuning the parameters during the training process, the Adam optimizer is the best.

[0113] ④ Gradient descent learning rate

[0114] Generally, the target value obtained from the initial parameters is relatively far from the global optimal value. As the number of iterations increases, it will get closer and closer to the global optimal value. The basic idea of learning rate decay is that the learning rate gradually decays as the training progresses, that is, a larger learning rate is used at the beginning to speed up the approach to the global optimal value, and a smaller learning rate is used later to improve stability, avoid skipping the global optimal value due to too large a learning rate, and ensure convergence to the global optimum. The learning rate is reduced by a factor of gamma every stepsize iterations.

[0115] The learning rate decreases in a polynomial curve.

[0116]

[0117] The learning rate decreases as the number of iterations increases.

[0118] LR(t) = baselr * (1 + gamma * iter) power

[0119] (4.4) Determination of model structure and parameters

[0120] Table 1 Model parameters

[0121]

[0122] The model structures of the networks for periodic data and aperiodic data are as follows Figure 10 shown:

[0123] The PDW homologous matching network proposed by the present invention mainly consists of two major modules, namely the similarity calculation module and the classification calculation module. The similarity calculation module of the network consists of a three-layer residual self-attention structure, which combines the ideas of the self-attention mechanism and the residual unit. In the training process, simply stacking the number of self-attention mechanism layers often causes the network performance to stagnate, and even counterproductively leads to a decline in network performance. Therefore, the idea of the residual neural network is introduced to retain a certain proportion of the input of the unit, and this proportion is also a network parameter that can be trained. This structure effectively solves the network degradation problems such as gradient explosion and gradient disappearance caused by the increase in the number of self-attention layers. The self-attention mechanism layer mainly calculates the Q, K, and V matrices through a convolutional layer with a convolution kernel size of 1×1, and calculates the similarity between Q and K using normalized dot product. After the data passes through the similarity calculation module of the network, an adaptive average pooling layer is provided to standardize its own dimension to facilitate the definition of the parameters of the subsequent network structure. After changing the data dimension, it is sent to the classification network module composed of a fully connected layer and a ReLU activation function to obtain the final matching result.

[0124] Performance analysis:

[0125] The variances of the measurement errors of the main matching parameters are changed respectively to test the performance of the PDW (pulse description word) homologous matching algorithm based on the self-attention mechanism proposed in this chapter. At the same time, the fuzzy clustering algorithm is used, and the self-attention mechanism network in the algorithm proposed in this chapter is replaced by CNN for comparison. Measurement errors obeying the normal distribution N(0, σ 2 ) are added to the carrier frequency, pulse width, and pulse repetition interval respectively. The unit of carrier frequency error is Hz; the unit of pulse width error is ns; the unit of pulse repetition interval error is ns.

[0126] As can be seen from Figure 11 , since the fuzzy clustering algorithm does not have the ability to match asynchronous data, even when the parameter measurement error is small, the matching correct rate is still much lower than that of the PDW (pulse description word) homologous matching algorithm based on the self-attention mechanism and the PDW (pulse description word) homologous matching algorithm based on CNN. Moreover, as the error increases, the decline rate of the matching correct rate of the fuzzy clustering algorithm is much larger than that of the other algorithms, and its tolerance to noise is lower. The matching performances of the PDW (pulse description word) homologous matching algorithm based on the self-attention mechanism and the PDW (pulse description word) homologous matching algorithm based on CNN both decline slowly as the variance of the measurement error increases. Since the self-attention mechanism is better at capturing the internal correlation of data or features, the matching correct rate using the self-attention mechanism is higher than that using CNN.

[0127] As can also be seen from Figure 11 , in this experiment, taking the residual unit as a variable, under the conditions that the training set and test set data are the same, the overall network structure remains unchanged, and the algorithm process remains unchanged, the matching performances of the network without the residual unit and the network with the residual unit are compared. As can be seen from the above three figures, the matching accuracy of the network with the residual unit added is significantly better than that of the network without the residual unit added.

[0128] It can also be known from the following table that the accuracy of traditional clustering for asynchronous data is far lower than that for synchronous data, while the algorithm proposed in the present invention has a matching accuracy of more than 95% for both synchronous data and asynchronous data.

[0129] Table 2 Matching correct rates of traditional algorithms and the algorithm of the present invention for different types of data

[0130]

[0131] In order to simultaneously meet the homologous matching requirements for synchronous cross-system data and asynchronous cross-system data and the homologous batching requirements for asynchronous same-system data, tests are carried out for synchronous cross-system data, asynchronous cross-system data, and asynchronous same-system data respectively under the conditions that the carrier frequency measurement error obeys N(0, 10000), the pulse width measurement error obeys N(0, 5), and the pulse repetition interval error obeys N(0, 5).

[0132] ① Synchronous cross-system data

[0133] When facing synchronous cross-system data, the homologous matching algorithm based on the self-attention mechanism performs excellently. The following table shows the matching accuracy rates of this algorithm for synchronous agile cross-system data and synchronous non-agile cross-system data. In this test, there are 10,000 groups for each type of homologous synchronous agile cross-system data, non-homologous synchronous agile cross-system data, homologous synchronous non-agile cross-system data, and non-homologous synchronous non-agile cross-system data.

[0134] Table 3 Matching accuracy rates for synchronous cross-system data

[0135]

[0136] ② Asynchronous cross-system data

[0137] When facing asynchronous cross-system data, the homologous matching algorithm based on the self-attention mechanism still maintains its matching performance. The following table shows the matching accuracy rates of this algorithm for synchronous cross-system data and asynchronous cross-system data. In this test, there are 10,000 groups for each type of homologous asynchronous agile cross-system data, non-homologous asynchronous agile cross-system data, homologous asynchronous non-agile cross-system data, and non-homologous asynchronous non-agile cross-system data.

[0138] Table 4 Matching accuracy rates for asynchronous cross-system data

[0139]

[0140] ③ Asynchronous same-system data

[0141] When facing asynchronous same-system data, the matching accuracy rate of the homologous matching algorithm based on the self-attention mechanism performs excellently. In this test, there are 10,000 groups for each type of homologous asynchronous agile same-system data, non-homologous asynchronous agile same-system data, homologous asynchronous non-agile same-system data, and non-homologous asynchronous non-agile same-system data.

[0142] Table 5 Matching accuracy rates for asynchronous same-system data

[0143]

Claims

1. A PDW homologous matching method based on the self-attention mechanism, characterized in that It includes the following steps; Step 1: Construct a dataset, and divide the dataset into the following three parts according to the division of matching scenarios: synchronous cross-system data, asynchronous cross-system data, and asynchronous same-system data; Step 2: Data preprocessing, preprocess the data in Step 1 before entering the algorithm; Step 3: Type judgment, respectively perform type judgment on the matching parameters of the data, carrier frequency RF, pulse width PW, and pulse PRI; the data types include non-agile, periodic agile, and non-periodic agile; If the data types of two groups of RF or two groups of PW or two groups of PRI are different, it is directly judged as non-homologous. If both groups of RF or both groups of PW or both groups of PRI are non-agile, the Euclidean distance is directly calculated to judge whether they are homologous. If both groups of RF or both groups of PW or both groups of PRI are periodic agile or non-periodic agile, they are sent to the self-attention mechanism network constructed in Step 4 to judge whether they are homologous; Step 4: Construct a self-attention mechanism network to judge whether two groups of RF or PW or PRI with carrier frequency RF, pulse width PW, and pulse PRI all being periodic agile or non-periodic agile are homologous.

2. The PDW homologous matching method based on the self-attention mechanism according to claim 1, wherein In the scenario of synchronizing cross-system data in step 1, the position of the target signal at almost the same moment is intercepted by two subsystems. At this time, the pulses from the two subsystems will partially or completely overlap. At site A i The received pulse time difference window contains site B j The received pulses; Synchronous cross-system data consists of homologous agile synchronous cross-system data, homologous non-agile synchronous cross-system data, non-homologous agile synchronous cross-system data, and non-homologous non-agile synchronous cross-system data; In the construction scenario of asynchronous cross-system data, the target signal is intercepted by two subsystems successively at similar but different time positions. At this time, the pulses from the two subsystems are difficult to align. At site A i There is basically no pulse received from site B within the pulse time difference window j The asynchronous cross-system data consists of homologous fast-varying asynchronous cross-system data, homologous non-fast-varying asynchronous cross-system data, non-homologous fast-varying asynchronous cross-system data, and non-homologous non-fast-varying asynchronous cross-system data; In the construction scenario of asynchronous same-system data, the dataset comes from PDW data received by the same subsystem at different times. After positioning is completed, the PDW parameter information of this time is used to match the PDW parameter information of the known target batch number within the same subsystem, so as to complete the function of target batch numbering; asynchronous same-system data consists of homologous agile asynchronous same-system data, homologous non-agile asynchronous same-system data, non-homologous agile asynchronous same-system data, and non-homologous non-agile asynchronous same-system data.

3. The PDW homologous matching method based on the self-attention mechanism according to claim 1, wherein The specific content of Step 2 is as follows: To extract the inter-pulse modulation information of the monitoring target, the pulse repetition interval PRI is calculated from the TOA measured by the system monitoring, and the calculation method is as follows: PRI(k) = TOA(k) - TOA(k - 1), k ∈ [2, N] where N is the number of pulses, PRI is the pulse repetition interval, and TOA is the time of arrival; After calculating PRI from the two groups of radar pulse description word data measured by the system monitoring respectively, the carrier frequency, pulse width, and pulse repetition interval of the two groups of PDW parameters are extracted and combined into three groups of data with a dimension of N×2. N is the number of pulses, and the forms of the three parameter matrices are shown in the following formula: Among them, taking RF 1n as an example, the subscript 1 indicates that this parameter comes from the first group, and the subscript n is the carrier frequency value of the nth pulse of this group of parameters.

4. The PDW homologous matching method based on the self-attention mechanism according to claim 1, characterized in that In Step 3, if both groups of RF or both groups of PW or both groups of PRI are not agile, the Euclidean distance is directly compared to judge whether they are the same; if the two groups of parameter types are different, it is directly determined as different; if both groups of parameters are determined to be periodic agile, they are sent to the network for periodic agile data to match whether they are the same; if both groups of parameters are determined to be non-periodic agile, they are sent to the network for non-periodic agile data to match whether they are the same. The three main matching parameters RF, PW, and PRI all go through the above process. If all three parameters can be matched to be the same, it is judged as homologous, otherwise it is judged as non-homologous.

5. The PDW homologous matching method based on the self-attention mechanism according to claim 1, wherein In step 3, type judgments are respectively performed on the matching parameters of the data, namely the carrier frequencies (RF1, RF2), pulse widths (PW1, PW2), and pulses (PRI1, PRI2). Here, the subscript 1 and subscript 2 represent that the parameters are from two different stations.

6. The PDW homology matching method based on the self-attention mechanism according to claim 1, characterized in that In step 3, the specific judgment of agility and periodicity is as follows: (3.1) Agility judgment Before judging whether the parameters are agile, first perform Min-max normalization on the carrier frequency RF, pulse width PW, and pulse PRI, and map the values to between [0, 1]. The formula is as follows: where x is a parameter sequence, min(x) and max(x) are the minimum and maximum values of the parameter sequence x respectively, and x * is the value after normalizing the parameter sequence x; After normalization, calculate the standard deviation of the parameters, and use the standard deviation to measure whether the data parameters are agile; (3.2) Periodicity judgment; Judge through the autocorrelation result. The autocorrelation formula is as follows: where x(n) is the parameter sequence, x(n - m) is the value of x(n) after being shifted by m, N is the number of parameters in the parameter sequence, and φ x (m) is the autocorrelation result; Since the time-domain autocorrelation is equivalent to the frequency-domain conjugate multiplication, and under the same amount of data, the operation speed of the frequency-domain conjugate multiplication is much higher than that of the time-domain autocorrelation. Therefore, the time-domain and frequency-domain conjugate multiplication is used to calculate the autocorrelation. The formula is as follows: φ x (m) = ifft(fft(x(n)) * conj(fft(x(n)))) where x(n) is the parameter sequence, fft is the fast Fourier transform, ifft is the inverse fast Fourier transform, conj is the complex conjugate operation, and φ x (m) is the autocorrelation result.

7. The PDW homologous matching method based on the self-attention mechanism according to claim 1, wherein In step 4, the process of sending the periodic and aperiodic agile data into the network to obtain the homologous result is as follows: First, perform Z_score standardization on the carrier frequency, pulse width, or pulse repetition interval of two groups of PDW in the parameter data. The formula is as follows: Among them, μ is the mean of all sample data, σ is the standard deviation of all sample data, and x is the carrier frequency (RF1, RF2) or pulse width (PW1, PW2) or pulse repetition interval (PRI1, PRI2); Then, send the standardized data into the pre-trained self-attention mechanism network. The self-attention mechanism network mainly includes the following two parts: a similarity calculation module and a classification module; the data first passes through the similarity calculation module composed of three layers of self-attention mechanisms. The self-attention mechanism can measure the similarity between the input sample itself and itself through the operations of the input Q, K, and V matrices. Taking the input data as the carrier frequency as an example, at this time, the similarity calculation module is calculating the similarity between (RF1, RF2); after obtaining the similarity matrix, send it into the classification module. The classification module is composed of a fully connected layer and an activation function. Judging whether they are homologous can essentially be regarded as a classification problem, and a homologous matching result of the parameters is obtained based on the similarity matrix.

Images