Radar pulse stream sorting method based on wide range embedding and mask interleaving

Abstract

A radar pulse stream sorting method based on wide-range embedding and mask interleaving includes the following steps: Step 1: preprocessing the target template time series signal; segmenting the time series data through a sliding window; Step 2: processing the data to generate simulated data, upgrading the data from the original multivariate time series to a wide-range embedding space as the input of the network; Step 3: masking the wide-range embedding space generated after conversion using the wide-range embedding method with a probability; Step 4: determining the loss function, optimizer, and learning strategy to obtain a network model; Step 5: inferring the model and adding a correction algorithm to the output result, performing sliding window sequence segmentation on the multivariate time series signal data composed of interleaved radar pulse descriptors in the inference stage, and serving as the input of the model inference stage. The present invention realizes high-granularity, sample-level pulse labeling and sorting for high-density interleaved radar pulse sequences.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar pulse stream sorting, and in particular relates to a radar pulse stream sorting method based on wide value range embedding and mask interleaving. Background Art

[0002] Radar emitter sorting has long been a key research topic in the field of electronic reconnaissance. Despite a surge in research achievements related to radar emitter signal sorting, overall, radar emitter signal sorting still faces shortcomings and requires overcoming existing bottlenecks to meet the demands of increasingly complex electromagnetic environments. Existing radar sorting algorithms are generally immature, with limited practical applications. They also generally fail to adequately account for non-ideal conditions such as high pulse density and severe pulse overlap. Existing methods generally assume that pulse signals are fully detected and the signal-to-noise ratio remains high. However, reality is far from ideal, and non-ideal conditions are common. On the one hand, electromagnetic interference often results in a low signal-to-noise ratio. On the other hand, the highly congested electromagnetic environment makes high-density radar pulses highly susceptible to high pulse overlap and severe pulse loss at the receiver end. These non-ideal conditions result in the measured performance of existing models being far below their theoretical sorting capabilities, significantly limiting the practical application of many models.

[0003] In the paper [Z.-M. Liu and PS Yu, "Classification, Denoising, and Deinterleaving of Pulse Streams With Recurrent Neural Networks," in IEEE Transactions on Aerospace and Electronic Systems, vol. 55, no. 4, pp. 1624-1639, August 2019, doi:10.1109 / TAES.2018.2874139], the authors used an RNN to deinterleave radar pulse streams, addressing the deinterleaving problem in pulse stream sequence patterns. However, this approach suffers from several drawbacks: It cannot automatically construct difficult sample cases, such as high-density overlap and pulse loss, making it difficult to improve the model's robustness. It also lacks specificity and fails to fully account for low signal-to-noise ratios and severe interference in complex electromagnetic environments. It can only deinterleave completely detected pulse sequences, failing to achieve highly granular, sample-level pulse labeling and sorting.

[0004] In the paper [Z.-M. Liu, "Pulse Deinterleaving for Multifunction Radars With Hierarchical Deep Neural Networks," in IEEE Transactions on Aerospace and Electronic Systems, vol. 57, no. 6, pp. 3585-3599, Dec. 2021, doi: 10.1109 / TAES.2021.3079571], the authors used a hierarchical deep learning model to solve the problem of different pulse groups belonging to the same radar. However, this method lacks an effective feature extraction mechanism and cannot fully exploit the deep features of multivariate time series data.

[0005] In the paper [X. Li, Z. Liu and Z. Huang, "Deinterleaving of Pulse Streams With Denoising Autoencoders," in IEEE Transactions on Aerospace and Electronic Systems, vol. 56, no. 6, pp. 4767-4778, Dec. 2020, doi: 10.1109 / TAES.2020.3004208], the authors used a denoising autoencoder to effectively deinterleave radar pulse streams. However, the algorithm's performance was limited, focusing primarily on denoising and reconstructing pulse streams. It failed to adequately address complex situations such as high-density overlap and low signal-to-noise ratio. It also lacked a mechanism for automatically constructing difficult samples, resulting in idealized training data and limited generalization capabilities.

[0006] In the field of radar pulse signal sorting, traditional methods still have limited performance and adaptability to the massive number of targets and dynamic environments in complex electromagnetic environments. Referenced deep learning-based algorithms all cluster radar pulse streams based on one-dimensional pulse descriptor data, which places higher performance demands on deep learning models. Therefore, the use of multivariate time-series pulse descriptor data to describe radar pulse signals, along with code embedding methods to obtain higher-dimensional feature descriptions, has attracted the attention of scholars in the field, making radar signal characteristics more distinct.

[0007] Currently, various research fields have begun to use deep learning algorithms to solve problems. However, current research results rarely use the wide-range embedding and mask interleaving construction methods combined with deep learning to solve signal sorting and recognition in complex electromagnetic environments.

[0008] However, today's electromagnetic environment and radar system are becoming increasingly complex, which puts forward requirements for research algorithms to be refined, intelligent and adaptive, making conventional reconnaissance methods face severe challenges.

[0009] Therefore, there is an urgent need for a method to quickly and accurately extract and identify target signals of interest from complex aliased signals. Summary of the Invention

[0010] To overcome the shortcomings of the above-mentioned prior art, the present invention aims to provide a radar pulse stream sorting method based on wide-range embedding and mask interleaving. This method can achieve high-granularity, sample-level pulse labeling and sorting for high-density interleaved radar pulse sequences. This method can cope with radar pulse sorting tasks under non-ideal conditions such as high density, severe pulse overlap, pulse loss, and low signal-to-noise ratio. To achieve the above-mentioned purpose, the technical solution adopted by the present invention is as follows:

[0011] A radar pulse stream sorting method based on wide range embedding and mask interleaving includes the following steps:

[0012] Step 1: In the training phase, preprocess the target template timing signal;

[0013] The time series signal is a non-interleaved multivariate time series signal composed of radar pulse description words. The non-interleaved multivariate time series signal is enhanced by combining the time scale transformation method, data mixing is achieved in parallel, and the time series data is segmented through a sliding window.

[0014] Step 2: Process the non-interleaved data after pre-processing and enhancement in step 1 to generate simulated data, where the simulated data simulates the pulse aliasing and interleaving situation in the radar pulse sorting problem;

[0015] In step 2, a wide-range scale interleaved signal embedding method is used to perform wide-range embedding processing on each dimensional time series contained in the simulated interleaved data generated after data enhancement, thereby upgrading the data from the original multivariate time series to the wide-range embedding space as the input of the network.

[0016] Step 3: Based on the characteristic that the original signal is interwoven with a small probability in the value range, the parameters of the mask function are calculated and set using the prior knowledge contained in the original signal. The wide-range embedding space generated after the conversion using the wide-range embedding method in step 2 is masked with a probability.

[0017] The aforementioned wide-range embedding method encodes and embeds different scales on the range through a family of positive integer periodic functions, thereby simultaneously embedding multivariate time series data into a wide-range space, and then extracts periodic patterns on the range and feature representations at different scales through convolution.

[0018] Compared to conventional embedding methods, the embedding space incorporates and preserves features of different scales. These features include information at different levels of interweaving, and the level of the current dimension also implies changes in the degree of interweaving, from non-trivial features (such as statistical features) to trivial features (deep features).

[0019] On this basis, masking the wide-range embedding space containing multiple scales can further construct more difficult samples; thus forcing the backbone neural network to classify signals based on deeper features rather than superficial simple statistical features. Following certain mask design principles to build a masking strategy can enhance the robustness and generalization ability of the model.

[0020] Step 4: Determine the number of backbone layers, loss function, optimizer, and learning strategy based on the performance and accuracy requirements of the sorting task, and ultimately obtain a network model for model inference in Step 5.

[0021] Step 5: In the model inference stage, the model obtained in step 4 is inferred, and a correction algorithm is added to the output result. The multivariate time series signal data composed of the interleaved radar pulse descriptors in the inference stage is segmented into sliding window sequences as the input of the model inference stage.

[0022] Due to limitations in video memory and other factors, the model's field of view is limited. Therefore, the following post-processing steps can utilize information beyond the model's field of view obtained in step 4: For the model's output annotation results containing noise, the output noise is filtered using the density anomaly filtering method in signal anomaly detection, thereby obtaining a more robust final prediction result. Optionally, step 1 specifically includes:

[0023] Step 1.1: Perform data enhancement on the non-interleaved signal by combining the time scale transformation method and implement data mixing in parallel:

[0024] Randomly sample from the training data and extract the data frame as the original data D0. Suppose that a L cut The time window of length D is used as the new data scaled , randomly discard part of the signal in proportion, randomly scale the time TOA, the scaling ratio obeys the normal distribution with mean μ and standard deviation σ, and limits the result to the simulated frequency shift limit α to β, simulating the impact of frequency shift; randomly shift the TOA time, the time shift amount obeys the exponential distribution, λ = 1, and limits the result to 0 to L cut The new TOA time is shifted to make its minimum value 0, and the offset of the original data TOA is removed;

[0025] Step 1.2: Let TAG_LEN be the number of target radar types to be classified. Traverse each tag index i∈{1,2,…,TAG_LEN} and measure the different types of features in the descriptor and the parameters of the pulse's time of arrival (TOA), radio frequency (RF), pulse width (PW), pulse amplitude (PA), and direction of arrival (DOA) according to the index.

[0026] Furthermore, TOA, RF, PW, PA, and DOA in step 1.2 above are the five most typical instantaneous parameters of the pulse descriptive word (PDW), that is, parameters that can be directly obtained through measurement. Different random transformations are performed on these parameters, including random scaling, random offset, and normalization operations.

[0027] Transforming RF features includes: extracting RF data corresponding to the current tag i right Perform normalization: in and They are The mean and standard deviation of Perform random scaling to obtain transformed data Where α~U(a,b) represents the uniform distribution U(a,b) with a superimposed random offset:

[0028] The transformation of PW features includes:

[0029] Extract the PW data corresponding to the current tag i right Perform normalization: in and They are The mean and standard deviation of Perform random scaling to obtain transformed data Where α'~U(a',b') represents the uniform distribution U(a',b') with a superimposed random offset:

[0030] Transformation of DOA features includes: superimposing random offset:

[0031] After the above enhancement processing, the overall offset is superimposed on the above features:

[0032]

[0033] ε RF , ε PW , ε DOA Indicates the random offset of RF, PW, and DOA as a whole, which obeys U(-a RF ,a RF )、U(-a PW ,a PW ) and U(-a DOA ,a DOA );

[0034] TOA is the arrival time, TAG is the tag, PA is the pulse amplitude, and no transformation is performed.

[0035] Step 1.3: Arrange the data in ascending order according to the TOA feature, intercept a time window, and perform the same random transformation enhancement on the data window. The reshaped data is the mixed data of the radar pulse signal in the interleaved state. Set the number of reshaping times and repeat the above interception sampling and transformation enhancement steps to generate more complex mixed data. At the same time, the sliding window is used to segment the sequence when mixing the window.

[0036] These data enhancement steps are designed to manually simulate the interleaved signals in real situations. The goal of the enhancement is to make the radar pulse interleaved signals similar to those actually received by the domain target.

[0037] Optionally, step 2 specifically includes:

[0038] Step 2.1: Using the wide-range scale interleaved signal embedding method, perform wide-range embedding on each dimension of the simulated interleaved data generated after data augmentation. This up-scales the data from the original multivariate time series into the wide-range embedding space and serves as the network input.

[0039] Assume that the length of N-dimensional variable s is L PDW ∈R L×N is the input signal, D is the dimension of the desired embedding, E i ∈R D is the embedding, which is determined by the input of the backbone network used, and the step size d is set step ∈N + (hereinafter abbreviated as d) and the scaling factor k∈N + To control the maximum modulus upper bound M = k D / d , on the contrary, d is determined by M, k step =D·log M k;

[0040] M is the maximum modulus upper bound, the maximum modulus is mmax =M / k 1 , which means the modulus m of the input data value max No information loss. Here, 1 is the default minimum modulus. Therefore, the input data needs to be linearly changed so that the data information can be reflected in the modulus 1 to m. max of the remaining systems.

[0041] Since M is the D-order exponential of k, we can select d and k to satisfy this condition. The wide range embedding E is obtained by the following formula PDW ∈R L×N×D :

[0042]

[0043] Where di+j∈{1,2,...,D} is an embedding dimension, x is the value of the input token, l is the input signal length, and n is the dimension of the input signal.

[0044] f ω＝1 (·) is a linear periodic function with a period of 1, expressed as follows:

[0045] f ω＝1 (x) = x mod 1·2-1

[0046] Wide range embedding is equivalent to transforming the original signal value through a family of linear periodic functions of positive integer multiples, and the wavelength forms a range from 1 to m max The positive integer k times of m contains information at a small scale only in the dimension with a larger m.

[0047] Optionally, step 3 specifically includes:

[0048] Step 3.1, mask design principles:

[0049] The mask function is used to further process the wide-range embedding space in step 3. The mask is represented as M, the fill value of the masked area is represented as z, and the masked embedding is represented as M*x+(1-M)*z. The following mask design principle is derived. For non-trivial f and trivial features g, M and z are required to satisfy:

[0050] ||f(M*x+(1-M)*z)-f(M'*x+(1-M')*z')|| 2 ≈||f(x1)-f(x2)|| 2

[0051] ||g(M*x+(1-M)*z)-g(M'*x+(1-M')*z')|| 2 >>||g(x1)-g(x2)|| 2

[0052] Step 3.2: Mask some of the dimensions embedded in step 2, and find the lowest dimensional function d for masking given the modulus m. low (·)for:

[0053] d low (m) = d·log k m

[0054] According to the above calculation results, d>d low By masking the dimension of (m), we can approximately construct an interleaving situation where the maximum and minimum differences are m signals;

[0055] The mask function is defined as F mask (·,d low ), the input is the wide range embedding tensor E in step 2 PDW ∈R L ×N× D , the output is

[0056] Step 3.3, select a suitable masking scheme. The following are three ways to mask on the time dimension. Suppose the input wide range embedding tensor is Here d low Represented as d low A random sample of (m);

[0057] The mean masking method is:

[0058]

[0059] The random value masking method is:

[0060]

[0061] Among them, R(·) means generating a tensor with the same elements uniformly distributed between [0,1];

[0062] The constant masking method is:

[0063]

[0064] Experiments have shown that the above methods all have similar and relatively good effects, but the random value mask has the strongest robustness.

[0065] Step 3.4, with a probability P hard For the input signal s PDW Masking any feature dimension of is called the difficulty probability of the hard sample. This is a hyperparameter that controls the probability of applying a mask to the input feature, thereby simulating the existence of a certain proportion of "hard samples".

[0066] Specifically, the difference Δ between the maximum and minimum values ​​of the original data PDW Determine the range to add the mask to:

[0067] {d∈(randint(d low (Δ PDW ),d low (2Δ PDW )),D)}

[0068] The function randint generates a discrete random integer variable uniformly distributed within [a, b]. low Represented as d low A random sample of (m).

[0069] The masking function determines the specific masking method, such as using mean masking or random masking. Using masked wide-range embedding as the training input of the model can improve the model's robustness to anomalies and noise.

[0070] At the same time, thanks to the wide range embedding generation, there is no need to normalize the enhanced data, only the signal meaningful information can be reflected in the modulus 1 to m max This range is wide enough to make a fixed affine transformation on the signal: a and b are determined by the distribution of such signals.

[0071] Optionally, step 4 specifically includes:

[0072] Step 4.1, the network includes a feature extraction backbone network and a feature decoder;

[0073] The input of the feature extraction backbone network is the pulse description word after wide range embedding and masking (optional), which then passes through the temporal depth separable convolution and convolutional feedforward network in sequence. The feature decoder processes the deep features output by the backbone network and outputs the classification confidence;

[0074] Among them, the depth-wise separable convolution is used to model temporal relationships and learn timing information. The convolutional feedforward network is composed of point-by-point convolution modules, which mix information between the feature dimension and the variable dimension. The first convolutional feedforward network is used to model channel relationships, and the second convolutional feedforward network is used to model variable relationships. Finally, the output feature dimension is connected to the MLP classifier. Then, based on the recognition results and sample labels, the parameters of the deep learning recognition model are continuously updated, and the model with the best performance on the validation set is saved as the final recognition model.

[0075] Specifically, the backbone network of the network model consists of a stack of time-series convolutional modules designed for radar time-series signals. Depthwise separable convolutions capture temporal relationships, while small convolution kernels capture local relationships. Finally, two convolutional feedforward modules provide nonlinear mapping. Each time-series convolutional module is responsible for capturing both the temporal and local spatial relationships of the input time-series signal.

[0076] Specifically, the module contains the following key parts:

[0077] 1. Depthwise separable convolution: This operation uses a decomposition of depthwise convolution and pointwise convolution to effectively capture patterns in the time domain while reducing computational complexity.

[0078] 2. Small kernel: In addition to the standard convolution kernel size (such as kernel_size = 51), a smaller convolution kernel (kernel_size = 5) is also used for convolution operation to capture more localized characteristic patterns in the time series signal;

[0079] 3. Batch normalization: Normalize the convolution output to accelerate convergence and improve model generalization ability;

[0080] 4. Convolutional Feedforward Module (ConvFFN): Contains two convolutional layers and GELU activation function, which introduces the necessary nonlinear mapping capabilities to the model;

[0081] In addition to the backbone network, the model also introduces a dual-head attention mechanism (DoubleAttention), which pays attention to the feature maps output by the module, improving the model's focus on important features. The attention module is located after the backbone network and before the classification head.

[0082] Step 4.2: Set the loss function to the multi-classification cross entropy function and use a 24-layer backbone network to effectively capture the long-term dependencies of time series signals. In addition, use the AdamW optimizer for training, adopt the learning rate decay (StepLR) and annealing strategy (CosineAnnealingLR), adjust the learning rate every certain epoch, and use gradient scaling (GradScaler) to achieve automatic / mixed precision training during training to improve the convergence speed.

[0083] Optionally, step 5 specifically includes:

[0084] Step 5.1: Segment the multivariate time series signal data composed of interleaved radar pulse descriptors in the test set into sliding window sequences and input them into the network for inference.

[0085] Step 5.2: Merge the windowed results of the network inference and perform sliding window sequence segmentation again. The length of this window is much longer than the training and inference length used by the model in step 4. Calculate the signal density and other characteristics of each sorted signal based on the windowed results, and then perform threshold difference maximum filtering on the signal density of continuous windows of a certain length.

[0086] Step 5.3: For the model that directly outputs a fixed-length signal category label sequence, individual signal category labels that are different from the current real signal may occasionally appear in the original output. These labels usually account for a very small proportion and are considered as noise points in the model output.

[0087] By using the density anomaly filtering method in signal anomaly detection, the following filtering conditions are designed to filter the noise points.

[0088] The following filter conditions are included (only the signal density feature is used as an example):

[0089] Filter condition 1: For each signal type i, calculate its corresponding signal density D signal , if the D of the current sequence segment signal If it is less than the corresponding threshold θ1, the softmax of the position is set to -1;

[0090] Filter condition 2: Set the softmax of all signal types whose proportion is less than the threshold θ2 to -1. In other words, if the prediction result of a certain category accounts for less than the threshold θ2, it will be regarded as a prediction noise;

[0091] Then all the softmax values ​​are set to -1, which is regarded as a noise signal;

[0092] In general, for the original output of the model, the categories with a signal density less than the threshold are first filtered out based on the signal density, and then the noise categories with a very small proportion are removed and supplemented with the main categories, so as to obtain a more robust final prediction result.

[0093] Beneficial effects of the present invention:

[0094] The present invention fully considers the urgent need for interleaved signal sorting in radar pulse sorting tasks, mines difficult samples through wide-range embedding generation and masking mechanisms, and compared with other single-pulse radar signal recognition methods, the present invention utilizes the characteristic that the input signal may contain important information at both large and small scales.

[0095] Different integer multiples of the modulus are applied to the original input data to extract features at different scales and represent them as wide-range embeddings. Wide-range embeddings are used to capture local patterns of small-probability interleaving of the original input signal across the range, thereby constructing interleaving patterns and increasing the proportion of difficult samples. This method combines innovative ideas such as Transformer absolute position encoding, "modulus normalization" technology, and temporal convolutional neural networks to fully utilize the information contained in the input signal at different scales and improve the model's feature representation capabilities.

[0096] Furthermore, by constructing interleaving patterns through masking, a targeted data augmentation strategy is implemented, helping the model learn the true data distribution and improving generalization and robustness. In theory, this method is equivalent to previous methods for difficult examples, but improves accuracy for easy examples, embodying the "easy first, difficult later" approach.

[0097] Compared with conventional data augmentation methods and network designs, this method does not require a large number of additional radar pulse signal sample data sets of other categories as mixing. Secondly, it greatly improves the performance of the model in the staggered radar pulse sorting scenario. BRIEF DESCRIPTION OF THE DRAWINGS

[0098] Figure 1 A schematic flow chart of a radar pulse stream sorting method based on wide-range embedding and mask interleaving provided in an embodiment of the present invention.

[0099] Figure 2 A schematic diagram of the network structure provided by an embodiment of the present invention.

[0100] Figure 3 This is a data enhancement flowchart provided by an embodiment of the present invention.

[0101] Figure 4 A flowchart for generating masked difficult samples provided by an embodiment of the present invention.

[0102] Figure 5 A comparison chart of changes in network training indicators in the ablation experiment provided by an embodiment of the present invention.

[0103] Figure 6 A comparison chart of the robustness test of the method of the present invention on the radar pulse sorting task provided by the embodiment of the present invention. DETAILED DESCRIPTION

[0104] The present invention will be described in further detail below with reference to the accompanying drawings.

[0105] A radar pulse stream sorting method based on wide range embedding and mask interleaving includes the following steps:

[0106] Step 1: Preprocess the multivariate time series signal of the non-interleaved radar pulse description word in the training phase, and combine Figure 3 Methods such as time scale transformation in

[15] are used to enhance data for non-interleaved signals and implement data mixing in parallel. Sequence segmentation is achieved through sliding windows.

[0107] Step 2: Considering that the value of the input signal may be important at both large and small scales, the present invention uses a wide-range scale interleaved signal embedding method to perform wide-range embedding on each dimension of the simulated data generated after data enhancement as the input of the network;

[0108] Step 3: Combine Figure 4 Based on the low-probability interleaving of the original signal across the range, we use the prior knowledge of the original signal to calculate and generate a masking function. We then mask the wide-range embeddings converted in step 2 with a probability, creating an interleaved scenario. This forces the backbone neural network to classify signals based on deep features of the sequence signal, rather than superficial, simple statistical features. By following certain mask design principles to construct a masking strategy, we enhance the model's robustness and generalization capabilities.

[0109] Step 4: Combine Figure 2 The network model structure is determined by determining the number of backbone layers, loss function, optimizer and learning strategy according to the performance and accuracy requirements of the sorting task, and finally obtaining the network model;

[0110] Step 5: Model inference and correction algorithm after output. The multivariate time series signal data composed of interleaved radar pulse descriptors from the inference phase is segmented into sliding window sequences and used as input for the model inference phase. The original sequence annotation results containing noise from the model output are filtered using the density anomaly filtering method used in signal anomaly detection, resulting in a more robust final prediction result.

[0111] Optionally, the multivariate time series signal of the radar pulse description word without interleaving in the training phase is preprocessed in step 1, and combined with Figure 3 Methods such as time scale transformation in

[15] are used to enhance data for non-interleaved signals and implement data mixing in parallel. Sequence segmentation is achieved through sliding windows, including:

[0112] Step 1.1: Data enhancement for non-interleaved signals, combined with Figure 3 Methods such as time scale transformation in

[15] are used to enhance data for non-interleaved signals and implement data mixing in parallel:

[0113] Randomly sample from the training data and extract the data frame as the original data D0. Suppose that a L cut The time window of length D is used as the new data scaled, randomly discard part of the signal in proportion, randomly scale the time TOA, the scaling ratio obeys the normal distribution with mean μ and standard deviation σ, and limits the result to the simulated frequency shift limit α to β, simulating the impact of frequency shift. Randomly shift the TOA time, the time shift amount obeys the exponential distribution, λ = 1, and limits the result to 0 to L cut The new TOA time is then shifted to its minimum value of 0, removing the offset from the original data TOA. Overall, the above steps simulate real-world conditions such as random window truncation, signal loss, frequency shift, and time shift, introducing a certain amount of data augmentation to the model and improving its generalization capabilities.

[0114] Step 1.2: Traverse each tag index i∈{1,2,...,TAG_LEN} and perform different random transformations on different types of features (RF, PW, DOA, TOA, TAG) in the description word according to the index, including random scaling, random offset, normalization and other operations.

[0115] Transforming RF features includes: extracting RF data corresponding to the current tag i right Normalize: in and They are The mean and standard deviation of Perform random scaling to obtain transformed data Where α~U(a,b) represents the uniform distribution U(a,b) with a superimposed random offset:

[0116] The transformation of PW features includes:

[0117] Extract the PW data corresponding to the current tag i right Normalize: in and They are The mean and standard deviation of Perform random scaling to obtain transformed data Where α'~U(a',b') represents the uniform distribution U(a',b') with a superimposed random offset:

[0118] Transformation of DOA features includes: superimposing random offset:

[0119] After the above enhancement processing, the overall offset is superimposed on the above features:

[0120]

[0121] ε RF , ε PW , ε DOA Indicates the random offset of RF, PW, and DOA as a whole, which obeys U(-a RF ,a RF )、U(-a PW ,a PW ) and U(-a DOA ,a DOA ).

[0122] In step 1.3, the data is sorted in ascending order according to the TOA feature. A time window is extracted and the same random transformation and enhancement are performed on the data window. The reshaped data is the mixed data of the interleaved radar pulse signals. By setting a certain number of reshapes and repeating the above sampling and transformation enhancement steps, more complex mixed data can be generated. Furthermore, a sliding window is used during the mixing window to segment the sequence.

[0123] These data enhancement steps are designed to manually simulate the interleaved signals in real situations. The goal of the enhancement is to make the radar pulse interleaved signals similar to those actually received by the domain target.

[0124] Optionally, in step 2, considering that the value of the input signal may be of great significance at both large and small scales, the present invention uses a wide-range scale interleaved signal embedding generation method to perform wide-range embedding on each dimension of the simulated data generated after data enhancement as the input of the network, including:

[0125] Step 2.1: Using the wide-range scale interleaved signal embedding method, perform wide-range embedding on the time series of each feature dimension of the simulated data generated after data augmentation and use it as the input of the network:

[0126] The present invention constructs wide range embedding by the following method, assuming s PDW ∈R L×N is the input signal, D is the dimension of the desired embedding, E i ∈R D is the embedding, which is determined by the input of the backbone network used, and the step size d is set step ∈N + (hereinafter abbreviated as d) and the scaling factor k∈N + To control the maximum modulus M = k D / d Conversely, d can also be determined by M and k. step =D·log M k.

[0127] M is the upper bound of the maximum modulus, and the maximum modulus is m max =M / k 1 , which means the modulus m of the input data value max No information loss. Here, 1 is the default minimum modulus. Therefore, the input data needs to be linearly changed so that the data information can be reflected in the modulus 1 to m. max The remaining series of

[0128] H(s PDW mod m max )≈H(s PDW )

[0129] H(s PDW mod 1)<<H(s PDW )

[0130] Since M is the D-order exponential of k, we can select d and k to satisfy this condition. The wide range embedding E is obtained by the following formula PDW ∈R L×N×D :

[0131]

[0132] where d step i+j is a dimension of embedding, and x is the value of the input token.

[0133] f ω＝1 (·) is a linear periodic function with a period of 1, and the expression is as follows:

[0134] f ω＝1 (x) = x mod 1·2-1

[0135] In general, wide-range embedding is equivalent to a family of linear periodic functions with a positive integer multiple, and the wavelength forms a positive integer k multiple from 1 to M, which can prevent information leakage in subsequent masks, that is, only the dimension with a larger m contains information at a small scale.

[0136] Optionally, step 3 is combined with Figure 4 Based on the low-probability interleaving of the original signal across the range, we use the prior knowledge of the original signal to calculate and generate a masking function. We then mask the wide-range embedding converted in step 2 with a probability, creating an interleaved state. This forces the backbone neural network to classify signals based on deep features of the sequence signal, rather than superficial, simple statistical features.

[0137] The aforementioned wide-range embedding method encodes and embeds different scales on the range through a family of positive integer periodic functions, thereby simultaneously embedding multivariate time series data into a wide-range space, and then extracts periodic patterns on the range and feature representations at different scales through convolution.

[0138] Compared to conventional embedding methods, the embedding space incorporates and preserves features of different scales. These features include information at different levels of interweaving, and the level of the current dimension also implies changes in the degree of interweaving, from non-trivial features (such as statistical features) to trivial features (deep features).

[0139] On this basis, masking the wide-range embedding space containing multiple scales can further construct more difficult samples; thus forcing the backbone neural network to classify signals based on deeper features rather than superficial simple statistical features. Following certain mask design principles to build a masking strategy can enhance the robustness and generalization ability of the model.

[0140] include:

[0141] Step 3.1, mask design principles:

[0142] The mask function is used to further process the dataset in step 3. Let the mask be M and the fill value of the masked area be z. The masked embedding can be expressed as M*x+(1-M)*z. Therefore, the following mask design principles are derived. For useful features f and trivial features g, M and z are required to satisfy:

[0143] ||f(M*x+(1-M)*z)-f(M'*x+(1-M')*z')|| 2 ≈||f(x1)-f(x2)|| 2

[0144] ||g(M*x+(1-M)*z)-g(M'*x+(1-M')*z')|| 2 >>||g(x1)-g(x2)|| 2

[0145] Step 3.2: Mask some of the dimensions embedded in step 2. Given the modulus m, find the lowest dimensional function d for masking low (·)for:

[0146] d low (m) = d·log k m

[0147] According to the above calculation results, d>d low By masking the dimension of (m), we can approximately construct an interleaving situation where the maximum and minimum differences are m signals.

[0148] The mask function is defined as F mask (·,d low ), the input is the embedding tensor E in step 2 PDW ∈R L×N× D , the output is

[0149] Step 3.3, select a suitable masking scheme. The following are three ways to mask in the time dimension. Here d low Represented as d low A random sample of (m).

[0150] The mean masking method is:

[0151]

[0152] The random value masking method is:

[0153]

[0154] Here, R(·) represents the generation of a tensor with the same elements uniformly distributed between [0,1].

[0155] The constant masking method is:

[0156]

[0157] Experiments have shown that the above methods all have similar and relatively good effects, but the random value mask has the strongest robustness.

[0158] Step 3.4, with a probability P hard For the input signal s PDW Masking any feature dimension of is called the difficulty probability of the hard sample. This is a hyperparameter that controls the probability of applying a mask to the input feature, thereby simulating the existence of a certain proportion of "hard samples".

[0159] Specifically, the difference Δ between the maximum and minimum values ​​of the original data PDW Determine the range to add the mask to:

[0160] {d∈(randint(d low (Δ PDW ),d low (2Δ PDW )),D)}

[0161] The function randint generates a discrete random integer variable uniformly distributed within [a, b]. low Represented as d low A random sample of (m).

[0162] The masking function determines the specific masking method, such as using mean masking or random masking. Masking training data can improve the robustness of the model to anomalies and noise.

[0163] At the same time, thanks to the wide range embedding generation, there is no need to normalize the enhanced data, only the signal meaningful information can be reflected in the modulus 1 to m max This range is wide enough to just do a fixed affine transformation on the signal: a and b are determined by the distribution of such signals.

[0164] Optionally, step 4 is combined with Figure 2 The network model structure is constructed by determining the number of backbone layers based on the hardware performance of the training platform, and then determining the loss function, optimizer, and learning strategy. Finally, the network model is obtained, including:

[0165] Step 4.1. The network includes a feature extraction backbone network and a feature decoder. The backbone network for feature extraction inputs the obtained pulse description words into the model, which passes through the time series depth separable convolution and convolutional feedforward network in sequence. Among them, the depth separable convolution is used to model time relationships and mainly learns time series information, while the convolutional feedforward network is composed of point-by-point convolution modules, which mix information between feature dimensions and variable dimensions. The first convolutional feedforward network is used to model channel relationships, and the second convolutional feedforward network is used to model variable relationships. Finally, the output feature dimension is connected to the MLP classifier, and then based on the recognition results and sample labels, the parameters of the deep learning recognition model are continuously updated, and the model with the best performance on the validation set is saved as the final recognition model.

[0166] In step 4.2, set the loss function to the multi-class cross entropy function. In this example, a 24-layer backbone network is used to effectively capture the long-term dependencies of time series signals. Furthermore, the AdamW optimizer is used for training, with a learning rate decay (StepLR) and an annealing strategy (CosineAnnealingLR) employed, adjusting the learning rate at regular epoch intervals. Gradient scaling (GradScaler) can also be used during training to achieve automatic / mixed-precision training and improve convergence speed.

[0167] Optionally, the correction algorithm after the model inference and output of step 5 is performed. The multivariate time series signal data composed of the interleaved radar pulse description words in the inference phase is segmented into a sliding window sequence as the input of the model inference phase.

[0168] Due to limitations on video memory and other factors, the model's field of view is limited. Therefore, the following post-processing steps can utilize information beyond the model's field of view obtained in step 4: For the noisy labeling results output by the model, the output noise is filtered through the density anomaly filtering method in signal anomaly detection, thereby obtaining a more robust final prediction result.

[0169] For the original sequence annotation results containing noise output by the model, the output noise is filtered out using the density anomaly filtering method in signal anomaly detection, thereby obtaining a more robust final prediction result, including:

[0170] Step 5.1: Segment the multivariate time series signal data composed of interleaved radar pulse descriptors in the test set into sliding window sequences and input them into the network for inference.

[0171] In step 5.2, combine the windowed results of the network inference and perform another sliding window sequence segmentation. This window length is much longer than the training and inference length used by the model in step 4. Based on the windowing, calculate the signal density and other characteristics of each sorted signal. Then, perform threshold difference maximum filtering on the signal density of continuous windows of a certain length.

[0172] Step 5.3: For models that directly output a fixed-length sequence of signal category labels, individual signal category labels that differ from the current true signal may occasionally appear in the original output. These labels usually account for a very small proportion and can be considered noise in the model output. Therefore, the present invention uses the density anomaly filtering method in signal anomaly detection to design the following filtering conditions to filter out these noises. The following filtering conditions are included (only the signal density feature is used as an example):

[0173] Filter condition 1: For each signal type i, calculate its corresponding signal density D signal If the D of the current sequence segment signal If it is less than the corresponding threshold θ1, the softmax of that position is set to -1.

[0174] Filter condition 2: Set all softmax values ​​of signal types whose proportion is less than the threshold θ2 to -1. In other words, if the proportion of the prediction result of a certain category is less than the threshold θ2, it will be regarded as a prediction noise.

[0175] Then all the softmax values ​​are set to -1 and regarded as noise signals.

[0176] In general, for the original output of the model, the categories with a signal density less than the threshold are first filtered out based on the signal density, and then the noise categories with a very small proportion are removed and supplemented with the main categories, so as to obtain a more robust final prediction result.

[0177] Example 1

[0178] 1. Experimental conditions:

[0179] To demonstrate the effectiveness of this method, this example validated the proposed method with real-world data from a competition. The training set for the experiment set the duration of each radar signal type to 10 seconds, with signal type labels ranging from 0 to 11. The training set contained sample data for 12 signal types. Each signal data point consisted of a series of time-series data points, each primarily described by six characteristic parameters, as shown in the table below. From left to right, they are: timestamp TOA, carrier frequency RF, pulse width PW, pulse amplitude PA, angle of arrival DOA, and tag ID. Each signal type contained only pure, noise-free signals of that type (free from interleaving, loss, errors, and interference).

[0180] PDW is often used for radar signal sorting. A typical PDW is composed of the following expression:

[0181]

[0182] Where PA is the pulse amplitude, TOA is the pulse arrival time, PW is the pulse width, RF is the carrier frequency, DOA is the pulse arrival angle, i is the radar pulse number received in chronological order, and N is the total number of intercepted pulses.

[0183] After the radar reconnaissance system receives a pulse signal, the parameter detection and measurement module first measures the pulse's key parameters, including time of arrival (TOA), radio frequency (RF), pulse width (PW), pulse amplitude (PA), and direction of arrival (DOA). The aforementioned TOA, RF, PW, PA, and DOA are the five most typical instantaneous parameters of the pulse descriptive word (PDW), i.e., parameters that can be directly obtained through measurement. In addition to instantaneous parameters, parameters that can be obtained through multiple measurements or calculations are secondary parameters, the most prominent of which is the pulse repetition interval (PRI). Whether instantaneous or secondary, they essentially belong to interpulse modulation characteristics.

[0184] Therefore, the data set examples in the verification process are as follows:

[0185] Table 1 Composition of description words in radar pulse dataset

[0186]

[0187]

[0188] Training data: interleaved data of multiple signal types. An interleaved radar pulse sequence data set is produced according to the method of step (2) of the invention. The total sample set is divided into a training set and a test set, where the training set accounts for 80% and the validation set accounts for 20%.

[0189] Test data: Interleaved data of various signal types. A total of three scenarios are set for verification data. Each scenario data is saved as a .txt file. The data format is the same as the training data.

[0190] The experiment consists of three test sets, each validating the radar sorting performance under several interleaved signal scenarios. The details are as follows:

[0191] Test sample 1: 4 interleaved signals, signal numbers 0, 2, 5, 7 type signal interleaved pulse sequence, a total of 999038 test samples as the test set for two target scenarios.

[0192] Test sample 2: 7 interleaved signals, signal numbers 0, 1, 2, 4, 5, 6, 7, 8 type of signal interleaved pulse sequence, a total of 1762620 test samples as the test set for two target scenarios.

[0193] Test sample 3: All 12 interleaved signals. A total of 2,579,053 test samples are used as the test set for the two target scenarios.

[0194] It's important to note that before data is input to the network, the time series sliding window must be split into a specified length. In this experiment, due to hardware memory limitations, the radar pulse sequence length is set to 1024 points. The training set contains radar pulse descriptor parameter information, and each sample has a dimension of (1024, 5).

[0195] Hardware platform: Server, CPU is Intel(R) Core(TM) i9-11900K, main frequency is 3.50GHz, GPU is NVIDIA GeForce RTX 3090;

[0196] Operating system: Ubuntu 20.04.6LTS;

[0197] Development tools: Python 3.7, torch 1.14.1, torchvision 0.13.0.

[0198] 2. Experimental content:

[0199] Regarding the aforementioned Figure 1 、 Figure 2 and Figure 3The technical solution, network structure and algorithm flow chart shown in the figure are described. The embodiment of the present invention further illustrates the technical performance and effect of the above technical solution through specific simulation experiments. The specific simulation conditions and parameters are as follows:

[0200] The embodiment of the present invention uses the Pytorch framework to build an alternative network model, uses NVIDIA GeForce GTX3090 to train the model, adopts AdamW optimizer to train the model parameters for updating, sets the batch size to 50, sets the initial learning rate to 0.004, and sets the network training to 400 rounds.

[0201] Under the above experimental conditions, the method of the present invention is used to pre-process all the non-interleaved radar pulse waveform data, extract the pulse description word parameters corresponding to each pulse in the data, and construct the original data sample set. Then, according to Figure 3 The training data is augmented using wide-range signal embedding generation and time-scale transformation, with parallel data mixing. The divided training and test data are then normalized, wide-range embedding encoding is introduced, and the sequence data is segmented using a sliding window to generate data windows for model training. A mechanism for masking the wide-range embeddings to construct interleaved patterns is also introduced within the training data set as a data augmentation measure to increase model robustness. This is used to mine difficult samples from radar pulse sequences under non-ideal conditions.

[0202] Then follow Figure 2 The network structure shown here develops a network model based on the fusion of multi-dimensional time series features. The backbone network consists of a stack of time series convolutional modules designed for radar time series signals. Depthwise separable convolutions capture temporal relationships, while small convolution kernels capture local relationships. Finally, two convolutional feedforward modules provide nonlinear mapping. Each time series convolutional module captures both the temporal and local spatial relationships of the input time series signal. Specifically, the module contains the following key components:

[0203] 1. Depthwise separable convolution: This operation uses a decomposition of depthwise convolution and pointwise convolution, which can effectively capture patterns in the time domain while reducing computational complexity.

[0204] 2. Small kernel: In addition to the standard convolution kernel size (such as kernel_size=7), a smaller convolution kernel (kernel_size=5) is also used for convolution operation to capture more local feature patterns in the time series signal.

[0205] 3. Batch normalization: Normalize the convolution output to accelerate convergence and improve model generalization ability.

[0206] 4. Convolutional feedforward module (ConvFFN): Contains two convolutional layers and GELU activation function, which introduces the necessary nonlinear mapping capabilities into the model.

[0207] In addition to the backbone network, the model also introduces a dual-head attention mechanism (DoubleAttention), which pays attention to the feature maps output by the module, improving the model's focus on important features. The attention module is located after the backbone network and before the classification head.

[0208] The network model takes multivariate time series data, such as radar signals, as input, which is mapped into a high-dimensional feature space via the Embedding layer. The backbone network then captures temporal and local spatial relationships, while the attention module focuses on important features. Finally, a linear classification head makes a classification prediction for the target. The network's forward propagation process involves the following steps: The input is position information pos, with a shape of [B, L, M, pos_D]. pos is transformed into [B, M, D, L] through permutation. The embedded features are then fed into ModernTCNBlock for processing. The DoubleAttention module is applied after the backbone. The processed features are flattened and passed through a linear layer for classification prediction.

[0209] Specific network model parameters include:

[0210] 1. Input parameters: M is the number of variables in the multivariate sequence; L is the length of the time window;

[0211] 2. Embedding layer parameters: P is the size of the embedding layer convolution kernel; S is the embedding layer step size; D is the embedding dimension.

[0212] 3. Backbone parameters: M: same as input M; same as embedding layer D; kernel_size is the size of the depthwise separable convolution kernel; r is the expansion rate of the ConvFFN module; num_classes: the number of categories for the classification task.

[0213] 4. Attention module parameters: DoubleAttention module parameters: D, M, L.

[0214] So a typical parameter configuration can be: M=5, L=1024, num_classes=12, D=128, P=1, S=1, kernel_size=7, r=2, num_layers=2.

[0215] These parameters need to be adjusted according to the actual task data, resource conditions and required classification accuracy.

[0216] The parameters used for wide range embedding are D=128, d=8, k=2, then M=65536. As long as the meaningful information of the signal is not lost in the remainder system modulo 1, 2, 4, ..., 65536, the algorithm conditions are met.

[0217] The network model is trained for 300 epochs using a large number of interleaved and masked radar pulse interleaved difficult sample signals. The indicators and loss function changes during the training process are as follows: Figure 4 shown.

[0218] The network backend classifier recognizes the modulated signals of highly interleaved radar pulse sequences and outputs the labeled classification results. Ultimately, a convergent and stable deep metric learning network model is trained, thereby improving the performance of radar pulse signal sorting and recognition.

[0219] 3. Results Analysis

[0220] After training the multivariate temporal feature fusion network model, Figure 1 The test set was processed using the same processing flow and the sorting effect was statistically analyzed. At the same time, several mainstream deep learning models were used to conduct comparative experiments on the same test set. The recognition effects of each model are shown in Table 2.

[0221] Table 2 Comparison of recognition indicators of mainstream deep learning algorithms

[0222]

[0223] The backbones of the selected comparison structures Transformer and ModernTCN in the table are the same as those used in this application;

[0224] Table 3. Comparison of ablation experiment indicators

[0225]

[0226] As can be seen from Table 3, the wide range embedding of this application can greatly improve the stability of the training process. By constructing interleaved samples through masking, the loss can be converged and the accuracy can be improved. Figure 5From the comparison chart of the changes in network training indicators in the provided ablation experiment, it can be seen that: using the wide-range embedding method of this application, the loss curve of the training process is smoother and more stable, and tends to converge at an earlier training stage, which shows that the embedding method can effectively improve the training stability of the model. After the introduction of masked construction of interleaved samples, although the loss curve still fluctuates greatly during the training stage, it can still converge well and reach a lower loss value in the end. This shows that masked interleaving can increase the difficulty of the samples and make the model more robust, thereby achieving better performance in subsequent training; finally, after introducing the method of this application, various indicators of network training such as accuracy, precision, recall, etc. have been further improved after using wide-range embedding and masked interleaving, further confirming the effectiveness of this method.

[0227] In general, the experimental results show that the wide-range embedding and mask interleaving technology proposed in this application can enhance the stability and robustness of model training, thereby achieving the purpose of improving the model's prediction accuracy and generalization ability.

[0228] Combine Figure 6 A comparison chart of the robustness test of the method of the present invention on the radar pulse sorting task is provided, which shows the robustness performance of four different methods on the radar pulse sorting task. The vertical axis represents the accuracy, and the horizontal axis represents the signal-to-noise ratio. The lower the signal-to-noise ratio, the worse the quality of the test data and the greater the degree of difficulty. Through the curve representing "wide range embedding + random mask (method of this application)", it can be seen that under conditions of lower signal-to-noise ratio (i.e., poor data quality), its robustness score decreases the slowest, maintaining a relatively high performance. The present invention adopts a method of wide range embedding and masking to construct interleaved samples, which can significantly improve the robustness and noise resistance of the model in complex environments, and maintain better sorting performance when processing radar pulse data of poor quality, verifying the superiority of this method in practical applications.

[0229] Compared with the existing technology, the present invention has the following significant advantages: (1) It realizes the high-grained sample-level sorting and labeling of interleaved radar pulse sequences: Unlike the existing methods that only target non-interleaved and ideal single pulse situations, the present invention can cope with complex non-ideal scenarios such as high density, severe pulse overlap, pulse loss, and low signal-to-noise ratio, and finely sort and label interleaved pulse sequences at the sample level, innovatively breaking through the limitations of traditional methods; (2) It proposes a new data enhancement method using wide-range embedding and mask interleaving: By using wide-range embedding to capture important signal features of large and small scales, and combining the mask function generated a priori from the original signal, it constructs interleaved data samples, so that the model not only learns statistical features, but also can classify from deep-level sequence features. This data enhancement method improves the robustness and generalization ability of the model; (3) It designs a complete end-to-end sorting network process: including preprocessing enhancement, sequence embedding, mask interleaving, network model training, inference and filter correction, forming a complete and effective interleaved pulse sorting solution that can effectively improve recognition accuracy and noise resistance. In summary, the present invention effectively solves the problem that existing methods are unable to handle complex interlaced scenes, and innovatively designs data enhancement and network architecture, which is of great significance to improving the performance of radar pulse sorting.

[0230] The technical points of the present invention are:

[0231] 1. Preprocess the multivariate time series signal composed of non-interleaved radar pulse descriptors during the training phase, perform data enhancement on the non-interleaved signal using a time scale transformation method, implement data mixing in parallel, and segment the sequence using a sliding window.

[0232] 2. Using the wide-range scale interleaved signal embedding method, a wide-range embedding is performed on each dimension of the simulated data generated after data augmentation, which is then used as the network input.

[0233] 3. Calculate and generate a mask function based on the prior knowledge contained in the original signal, mask the wide-range embedding converted in step 2 with a probability, construct an interleaving scenario, and build a masking strategy to enhance the robustness and generalization ability of the model;

[0234] 4. Build the network model structure, determine the number of backbone layers based on the performance and accuracy requirements of the sorting task, determine the loss function, optimizer, and learning strategy, and ultimately obtain the network model;

[0235] 5. Perform sliding window sequence segmentation on the multivariate time series signal data composed of interleaved radar pulse descriptors in the inference stage and use it as the input of the model inference stage. For the original sequence labeling results containing noise output by the model, the output noise is filtered through the density anomaly filtering method in signal anomaly detection to obtain the final prediction result.

Claims

1. A radar pulse stream sorting method based on wide range embedding and mask interleaving is characterized by: The following steps are included: Step 1: In the training phase, preprocess the target template time series signal; the time series signal is a non-interleaved multivariate time series signal composed of radar pulse descriptors, and the time series data is segmented by a sliding window; Step 2: Process the non-interleaved data after preprocessing and enhancement in step 1 to generate simulated data. The simulated data simulates the pulse aliasing and interleaving in the radar pulse sorting problem. The data is upgraded from the original multivariate time series to a wide range embedding space as the input of the network. Step 3: Based on the characteristic that the original signal is interwoven with a small probability in the value range, the parameters of the mask function are calculated and set using the prior knowledge contained in the original signal. The wide-range embedding space generated after the conversion using the wide-range embedding method in step 2 is masked with a probability. Step 4: Determine the number of backbone layers, loss function, optimizer, and learning strategy based on the performance and accuracy requirements of the sorting task, and ultimately obtain a network model for model inference in Step 5. Step 5: In the model inference stage, the model obtained in step 4 is inferred, and a correction algorithm is added to the output result. The multivariate time series signal data composed of the interleaved radar pulse descriptors in the inference stage is segmented into sliding window sequences as the input of the model inference stage.

2. The radar pulse stream sorting method based on wide range embedding and mask interleaving according to claim 1 is characterized in that: Step 1 specifically includes: Step 1.1: Perform data enhancement on the non-interleaved signal by combining the time scale transformation method and implement data mixing in parallel: Randomly sample from the training data and extract the data frame as the original data D0. Suppose that a L cut The time window of length D is used as the new data scaled , randomly discard part of the signal in proportion, randomly scale the time TOA, the scaling ratio obeys the normal distribution with mean μ and standard deviation σ, and limits the result to the simulated frequency shift limit α to β, simulating the impact of frequency shift; randomly shift the TOA time, the time shift amount obeys the exponential distribution, λ = 1, and limits the result to 0 to L cut The new TOA time is shifted to make its minimum value 0, and the offset of the original data TOA is removed; Step 1.2: Let TAG_LEN be the number of target radar types to be classified. Traverse each tag index i∈{1,2,…,TAG_LEN} and measure the parameters of the arrival time TOA, carrier frequency RF, pulse width PW, pulse amplitude PA, and direction of arrival DOA of different types of features in the descriptor according to the index. Step 1.3: Arrange the data in ascending order according to the TOA feature, intercept a time window, and perform the same random transformation enhancement on the data window as in steps 1.1 and 1.

2. The reshaped data is the mixed data of the radar pulse signal in the interleaved state. Set the number of reshaping times and repeat the above interception sampling and transformation enhancement steps to generate more complex mixed data. At the same time, the sliding window is used to segment the sequence in the mixed window.

3. The radar pulse stream sorting method based on wide range embedding and mask interleaving according to claim 2 is characterized in that: The TOA, RF, PW, PA, and DOA in step 1.2 above are the five instantaneous parameters of the pulse descriptor PDW, that is, the parameters directly obtained through measurement. Different random transformations are performed on these parameters, including random scaling, random offset, and normalization operations. Transforming RF features includes: extracting RF data corresponding to the current tag i right Perform normalization: in and They are The mean and standard deviation of Perform random scaling to obtain transformed data Where α~U(a,b) represents the uniform distribution U(a,b) with a superimposed random offset: The transformation of PW features includes: Extract the PW data corresponding to the current tag i right Perform normalization: in and They are The mean and standard deviation of Perform random scaling to obtain transformed data Where α'~U(a',b') represents the uniform distribution U(a',b') with a superimposed random offset: Transformation of DOA features includes: superimposing random offset: ε”~U(c”,d”); After the above enhancement processing, the overall offset is superimposed on the above features: ε RF , ε PW , ε DOA Indicates the random offset of RF, PW, and DOA as a whole, which obeys U(-a RF ,a RF )、U(-a PW ,a PW ) and U(-a DOA ,a DOA ); TOA is the arrival time, TAG is the tag, PA is the pulse amplitude, and no transformation is performed.

4. The radar pulse stream sorting method based on wide range embedding and mask interleaving according to claim 1 is characterized in that: Step 2 specifically includes: Step 2.1: Using the wide-range scale interleaved signal embedding method, perform wide-range embedding on each dimension of the simulated interleaved data generated after data augmentation. This up-scales the data from the original multivariate time series into the wide-range embedding space and serves as the network input. Assume that the length of N-dimensional variable s is L PDW ∈R L×N is the input signal, D is the dimension of the desired embedding, E i ∈R D is the embedding, which is determined by the input of the backbone network used, and the step size d is set step ∈N + , with the scaling factor k∈N + To control the maximum modulus upper bound M = k D / d , d step Abbreviated as d, conversely, d is determined by M and k step =D·log M k; M is the maximum modulus upper bound, the maximum modulus is m max =M / k 1 , which means the modulus m of the input data value max No information loss. Here, 1 is the default minimum modulus. Therefore, the input data needs to be linearly changed so that the data information can be reflected in the modulus 1 to m. max The remaining series of Since M is the D-order exponential of k, choose d and k to satisfy M = k D / d , the wide range embedding E is obtained by the following formula PDW ∈R L×N×D : Where di+j∈{1,2,...,D} is an embedding dimension, x is the value of the input token, l is the input signal length, and n is the dimension of the input signal; f ω＝1 (·) is a linear periodic function with a period of 1, expressed as follows: f ω＝1 (x)=x mod 1·2-1 Wide range embedding is equivalent to transforming the original signal value through a family of linear periodic functions of positive integer multiples, and the wavelength forms a range from 1 to m max The positive integer k times of m contains information at a small scale only in the dimension with a larger m.

5. The radar pulse stream sorting method based on wide range embedding and mask interleaving according to claim 4 is characterized in that: Step 3 specifically includes: Step 3.1, mask design principles: The mask function is used to further process the wide-range embedding space. The mask is represented as M, the fill value of the masked area is represented as z, and the masked embedding is represented as M*x+(1-M)*z. The following mask design principles are derived. For non-trivial f and trivial features g, M and z are required to satisfy: ‖f(M*x+(1-M)*z)-f(M′*x+(1-M′)*z′‖ 2 ≈‖f(x1)-f(x2)‖ 2 ||g(M*x+(1-M)*z)-g(M′*x+(1-M′)*z′)|| 2 >>||g(x1)-g(x2)|| 2 Step 3.2: Mask some of the dimensions embedded in step 2, and find the lowest dimensional function d for masking given the modulus m. low (·)for: d low (m)=d·log k m According to the above calculation results, d>d low By masking the dimension of (m), we can approximately construct an interleaving situation where the maximum and minimum differences are m signals; The mask function is defined as F mask (·,d low ), the input is the wide range embedding tensor E in step 2 PDW ∈R L×N×D , the output is Step 3.3, select a suitable masking scheme. The following are three ways to mask on the time dimension. Suppose the input wide range embedding tensor is d low Represented as d low A random sample of (m); The mean masking method is: The random value masking method is: Among them, R(·) means generating a tensor with the same elements uniformly distributed between [0,1]; The constant masking method is: Step 3.4, with a probability P hard For the input signal s PDW Any feature dimension of is used as a mask.

6. The radar pulse stream sorting method based on wide range embedding and mask interleaving according to claim 5 is characterized in that: The difference Δ between the maximum and minimum values ​​of the original data PDW Determine the range to add the mask to: {d∈(randint(d low (Δ PDW ),d low (2Δ PDW )),D)} The function randint generates a discrete random integer variable uniformly distributed within [a, b]. low Represented as d low A random sample of (m); The information that satisfies the signal is meaningful and is reflected in the modulo 1 to m max In the residual system of , a fixed affine transformation is performed on the signal: a and b are determined by the distribution of such signals.

7. The radar pulse stream sorting method based on wide range embedding and mask interleaving according to claim 1 is characterized in that: Step 4 specifically includes: Step 4.1, the network includes a feature extraction backbone network and a feature decoder; The input of the feature extraction backbone network is the pulse description word after wide-range embedding and masking, which then passes through the temporal depth-separable convolution and convolutional feedforward network in sequence. The feature decoder processes the deep features output by the backbone network and outputs the classification confidence. Among them, depthwise separable convolution is used to model temporal relationships and learn timing information. The convolutional feedforward network is composed of point-by-point convolution modules. The convolutional feedforward network mixes information between the feature dimension and the variable dimension. The first convolutional feedforward network is used to model channel relationships, and the second convolutional feedforward network is used to model variable relationships. Finally, the output feature dimension is connected to the MLP classifier. Then, based on the recognition results and sample labels, the parameters of the deep learning recognition model are continuously updated, and the model with the best performance on the validation set is saved as the final recognition model. Specifically, the backbone network of the network model consists of a stack of time-series convolution modules designed for radar time-series signals. It captures temporal relationships through depthwise separable convolutions and local relationships through small convolution kernels. Finally, two convolution feedforward modules provide nonlinear mapping. Each time-series convolution module is responsible for capturing the temporal and local spatial relationships of the input time-series signal. Step 4.2: Set the loss function to the multi-classification cross entropy function and use a 24-layer backbone network to effectively capture the long-term dependencies of time series signals. In addition, use the AdamW optimizer for training, adopt a learning rate decay and annealing strategy, adjust the learning rate every certain epoch, and use gradient scaling to achieve automatic / mixed precision training during training.

8. The radar pulse stream sorting method based on wide range embedding and mask interleaving according to claim 7 is characterized in that: Specifically, the temporal convolution module contains the following key parts: Depthwise separable convolution: uses a decomposition form of depthwise convolution and pointwise convolution; Small convolution kernel: In addition to the standard convolution kernel size kernel_size=51, a smaller convolution kernel kernel_size=5 is also used for convolution operation; Batch normalization: normalize the convolution output; Convolutional feedforward module: contains two convolutional layers and GELU activation function, which introduces the necessary nonlinear mapping capabilities into the model; The model also introduces a dual-head attention mechanism to pay attention to the feature map output by the module. The attention module is located after the backbone network and before the classification head.

9. The radar pulse stream sorting method based on wide range embedding and mask interleaving according to claim 1 is characterized in that: Step 5 specifically includes: Step 5.1: Segment the multivariate time series signal data composed of interleaved radar pulse descriptors in the test set into sliding window sequences and input them into the network for inference. Step 5.2: Merge the sliding windows of the network inference results and perform sliding window sequence segmentation again; the sliding window length is much longer than the training and inference length used by the model in step 4; calculate the signal density and other characteristics of each sorted signal based on the sliding window, and then perform threshold difference maximum filtering on the signal density of continuous windows of a certain length; Step 5.3: For models that directly output a fixed-length sequence of signal category labels, individual signal category labels that differ from the current true signal may occasionally appear in the original output. These labels usually account for a very small proportion and are considered noise in the model output. By using the density anomaly filtering method in signal anomaly detection, the following filtering conditions are designed to filter the noise points. Then all positions where the softmax values ​​are set to -1 are considered as noise signals; For the original output of the model, we first filter out categories with a signal density less than a threshold, then remove the noise categories with a very small proportion, and complement them with the main categories to obtain a more robust final prediction result.

10. The radar pulse stream sorting method based on wide range embedding and mask interleaving according to claim 9 is characterized in that: Contains the following filter conditions: Filter condition 1: For each signal type i, calculate its corresponding signal density D signal , if the D of the current sequence segment signal If it is less than the corresponding threshold θ1, the softmax of the position is set to -1; Filter condition 2: Set all softmax values ​​of signal types whose proportion is less than the threshold θ2 to -1. That is, if the proportion of the prediction results of a certain category is less than the threshold θ2, it will be regarded as a prediction noise.

Images