Radar signal increment sorting identification method and system based on pulse stream synthesis playback

Abstract

The invention discloses a radar signal increment sorting recognition method and system based on pulse stream synthesis playback, mainly solving the problem of poor increment sorting recognition effect in the prior art, and the implementation scheme comprises the following steps: obtaining a radar signal pulse description word sequence collected offline and a corresponding pulse-by-pulse radar tag; dividing the training data into a plurality of pure subsequences, adding the pure subsequences into a memory bank taking a category as a unit, extracting the pure subsequences from the memory bank according to a category probability, and performing time translation and interleaving synthesis to generate an aliasing playback sample; iteratively training the neural network by using the aliasing playback sample, and obtaining a trained incremental neural network through verification; a radar signal pulse description word sequence to be sorted and recognized is obtained in real time and input into the trained incremental neural network to obtain pulse-by-pulse radar tags, and deep feature scoring and stratified sampling results are obtained through various types of pure subsequences input into the network to maintain a memory bank. The method can improve the effect of incremental sorting recognition, is high in generalization ability, controllable in training cost and high in overall efficiency, and can be used for electronic reconnaissance in a dynamic signal environment.

Description

Technical Field

[0001] This invention belongs to the field of radar signal processing technology, and specifically relates to a radar signal incremental sorting and identification method and system, which can be used for electronic reconnaissance in dynamic signal environments. Background Technology

[0002] With the diversification of modern radar systems and the increasing congestion of the spectrum environment, electronic reconnaissance systems need to perform real-time sorting and identification of signals from various radar sources under constantly changing battlefield conditions. Existing deep learning methods typically train models offline on static, predefined training datasets and operate on fixed sets of categories. These methods experience significant performance degradation when faced with dynamic environments such as the emergence of new radar models, changes in signal patterns, and the disappearance and reappearance of old targets.

[0003] Some current research uses the concept of "incremental" processing to improve the online sorting performance of radar signals. For example, one type of method is based on streaming clustering, employing algorithms such as affinity propagation and Dirichlet process mixture models to achieve adaptive clustering of continuously arriving pulse data. Another type of method improves real-time performance by optimizing the objective function, such as sparse-constrained incremental nonnegative matrix factorization (INMF), and locally updating parameters as new data arrives. These methods primarily focus on data update efficiency and computational overhead, but remain within the traditional machine learning paradigm, failing to address the catastrophic forgetting phenomenon that occurs in deep neural networks during multi-stage training. Therefore, these methods struggle to maintain the sorting and recognition capabilities of deep neural networks in scenarios with dynamic categories or changing signal patterns.

[0004] In the field of deep learning, incremental learning has been extensively studied to enable models to retain old knowledge while continuously accepting new tasks. Typical incremental learning methods include three main categories: regularization constraint methods, dynamic network structure expansion methods, and data replay methods. Among them:

[0005] Regularization constraint methods introduce constraints into the objective function to limit the network's forgetting of old knowledge when learning new tasks. The core idea is to control the changes in parameters or outputs during the training process of new tasks within a reasonable range by designing appropriate regularization terms, thereby maintaining the network's ability to discriminate historical tasks. For example, the elastic weight solidification method uses the Fisher information matrix to characterize parameter importance and introduces a quadratic penalty in the loss function to suppress the shift of key parameters; forgetting-free learning (LwF) uses distillation loss to constrain the new network's prediction distribution of old task samples to be close to that of the old network. However, regularization constraint methods risk reducing adaptability to new tasks and are difficult to adapt to radar signal sorting and recognition scenarios where the distribution of data features varies greatly at different stages.

[0006] Dynamic network structure expansion methods address the catastrophic forgetting problem from the perspective of network capacity management. When a new task or category arrives, instead of forcibly compressing old and new knowledge within a fixed parameter space, they dynamically adjust the network structure to provide additional space for new knowledge or reallocate existing network resources, thereby reducing the risk of conflict between old and new knowledge. For example, Progressive Neural Networks (PNNs) fix existing network parameters for each new task and add parallel subnetworks while introducing lateral connections to facilitate knowledge transfer. Dynamically scalable networks (DENs) use sparse regularization to determine which old parameters can be reused and expand new units as necessary. However, dynamic network structure expansion methods also suffer from problems such as continuously increasing network size, higher inference overhead, and a high degree of dependence on task boundaries. Balancing the effectiveness of network structure expansion with network complexity remains a key direction for future research.

[0007] Data replay methods alleviate forgetting by explicitly reproducing old data or its approximations during a new training phase, and are the most intuitive and widely used approach. The basic idea is that when a new task or category arrives, the network is not only optimized on the new data, but samples from past tasks or pseudo-samples synthesized by the generative network are also incorporated into the training to maintain the network's discriminative ability on older tasks.

[0008] In their paper, "GDumb: A simple approach that questions our progress," Ameya Prabhu and other scholars... The paper "In Continual Learning" proposes a data replay method called GDumb, which directly stores the current sample with a finite amount of memory at each stage and trains it uniformly in the final stage. Mengyao Zhai et al., in their paper "Lifelong GAN: Continual Learning for Conditional Image Generation," proposed a data replay method called Lifelong GAN, which uses a generative adversarial network (GAN) to train a generator at each stage to approximate the historical data distribution, and then uses the generated pseudo-samples for replay.

[0009] Although the above data playback method is intuitive in principle, stable in performance, and has good operability in the context of radar signal sorting and identification applications, it lacks the ability to generalize to combinations of old and new signals in scenarios with dynamic category or signal pattern changes due to the lack of targeted design for radar signal sorting and identification problems, resulting in poor incremental sorting and identification effectiveness. Summary of the Invention

[0010] The purpose of this invention is to address the shortcomings of existing data playback methods by proposing a radar signal incremental sorting and identification method and system based on pulse stream synthesis playback. This method enables continuous learning and stable sorting and identification in dynamic environments, improves the generalization ability of new and old signal combinations under dynamic category or signal pattern change scenarios, and enhances the effectiveness of incremental sorting and identification.

[0011] The technical approach to achieving the objective of this invention is as follows: By synthesizing playback data based on online aliasing of pure pulse streams, aliasing sequences are dynamically generated during the training process, rather than relying on pre-stored fixed combinations, fundamentally improving the generalization ability of new and old signal combinations in scenarios with dynamic category or signal pattern changes; by introducing a memory management scheme based on class distribution perception and a sample distribution coverage enhancement strategy based on deep feature layer extraction, the balance of playback data at the category level and the coverage at the feature space level are ensured, reducing the ever-increasing sample storage cost and training time during long-term detection, and achieving a trade-off between training cost and sorting and recognition effectiveness.

[0012] Based on the above ideas, the technical solution of the present invention includes the following:

[0013] 1. A radar signal incremental sorting and identification method based on pulse stream synthesis and playback, characterized in that it includes:

[0014] (1) Obtain the radar signal pulse descriptor sequence collected offline, label it with radar tags, and divide it into training data and validation data;

[0015] (2) Each pulse descriptor sequence in the training data is split into multiple pure subsequences according to the category label;

[0016] (3) Add the pure subsequence to the class-based memory and perform class probability updates;

[0017] (4) Extract pure subsequences of multiple categories from the memory bank according to the category probability, perform time shifting and interleaving synthesis to generate aliased playback samples;

[0018] (5) Input the aliased playback samples into the neural network for iterative training, and verify them with the verification data to obtain the trained incremental neural network;

[0019] (6) The radar signal pulse descriptor sequence to be sorted and identified is acquired in real time, and it is used as test data to be input into the trained incremental neural network. The output representation corresponding to each pulse is output, and the radar category corresponding to the maximum value is selected as the radar label of the pulse.

[0020] (7) For a certain category of pure subsequences whose number in the memory bank exceeds the upper limit of the memory bank capacity, they are also input into the trained incremental neural network, output the deep features of the pure subsequences, score the deep features and perform stratified sampling, and update the memory bank capacity according to the stratified sampling results.

[0021] (8) Repeat steps (1) to (7).

[0022] Furthermore, the class probability update performed in (3) is to adjust the class probability based on the number of pure subsequences of each class.

[0023] Furthermore, in step (4), the pure subsequences of multiple categories extracted from the memory are time-shifted and interleaved to generate aliased playback samples, which includes:

[0024] 4a) Time-shift the extracted pure subsequence to obtain the time-shifted pulse arrival time sequence. :

[0025] 4b) First, merge the pure subsequences of multiple categories into one sequence. Then, sort all the pulses in the merged sequence according to the pulse arrival time. Finally, trim the sorted sequence to retain only the time intervals of the interleaved sequences from multiple radar categories to obtain the aliased playback sample.

[0026] Furthermore, in step (7), the scoring of deep features and stratified sampling, and the updating of the memory capacity based on the stratified sampling results, are implemented as follows:

[0027] 7a) Input each pure subsequence in the memory bank into the trained incremental neural network to obtain deep features;

[0028] 7b) Average each deep feature along the pulse dimension to obtain the feature representation of each pure subsequence;

[0029] 7c) Select a set of reference sequences and calculate the mean of the feature representations of all sequences in each set of reference sequences to obtain the centroid of the reference feature for that class;

[0030] 7d) Calculate the Euclidean distance between the feature representation of each pure subsequence and the reference feature centroid of the class to which the subsequence belongs, and use it as the representative score of the pure subsequence;

[0031] 7e) Divide the newly added pure subsequences of each category into multiple intervals according to their representative scores, and sample each interval separately, so that the proportion of the pure subsequences of that category in the memory bank that are newly added pure subsequences in this stage is %. :

[0032] 7g) Divide the historical pure subsequences of each category into multiple intervals according to their representative scores, and sample from each interval to ensure that the proportion of historical pure subsequences in the memory bank for that category is [percentage missing]. .

[0033] 2. A radar signal incremental sorting and identification device based on pulse stream synthesis and playback, characterized in that it comprises:

[0034] The data acquisition module is used to acquire offline collected pulse descriptor sequences and perform label annotation and training / validation division;

[0035] The pulse separation module is used to separate aliased pulses according to category labels and generate pure subsequences for each category.

[0036] The memory management module is used to add pure subsequences to the memory and perform class probability updates;

[0037] The playback sample generation module is used to extract pure subsequences from the memory bank based on class probability and perform time shifting and interleaving to synthesize aliased playback samples.

[0038] The network training module is used to input aliased playback samples into the neural network for iterative training and to use validation data for validation, so as to obtain the incremental neural network after training.

[0039] The online sorting and recognition module is used to input real-time test data into a trained incremental neural network and output pulse labels;

[0040] The memory maintenance module is used for scoring based on deep features, hierarchical sampling, and updating the memory capacity.

[0041] Compared with the prior art, the present invention has the following advantages:

[0042] 1. This invention overcomes the limitations of incomplete data combination in traditional data playback methods by generating aliased playback samples based on pure pulse streams, improves the generalization ability of new and old signal combinations in dynamic category or signal pattern change scenarios, and enhances the effectiveness of incremental sorting and recognition.

[0043] 2. This invention, through category probability control and deep feature hierarchical sampling memory management and maintenance, can ensure the balance of playback data at the category level and the coverage at the feature space level. While ensuring the effectiveness of sorting and recognition, it can reduce sample storage costs and training time, improve incremental sorting and recognition efficiency, and enhance engineering usability. Attached Figure Description

[0044] Figure 1 This is a flowchart illustrating the implementation of the radar signal incremental sorting and identification method based on pulse stream synthesis and playback according to the present invention.

[0045] Figure 2 This is a sub-flowchart of deep feature scoring and stratified sampling in the method of this invention;

[0046] Figure 3 This is a block diagram of the radar signal incremental sorting and identification device based on pulse stream synthesis and playback according to the present invention;

[0047] Figure 4 The graph shows the comparison results of different indicators for incremental sorting and identification using the present invention and existing joint training methods, fine-tuning methods, GDumb methods, and LwF methods.

[0048] Figure 5 This is a comparison chart showing the average training time of each stage of incremental sorting and identification using the present invention and existing joint training methods, fine-tuning methods, GDumb methods, and LwF methods. Detailed Implementation

[0049] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, other embodiments obtained by those skilled in the art without creative effort should all fall within the protection scope of the present invention.

[0050] It should be noted that the step numbers in the specification and claims of this invention are only for the purpose of clearly describing the embodiments of this invention and facilitating understanding, and their order is not limited.

[0051] Example 1: Radar signal incremental sorting and identification method based on pulse stream synthesis and playback.

[0052] In real-world, complex electromagnetic environments, the receiver continuously intercepts pulse signals from different radars. As the mission progresses, the system acquires new pulse data over multiple time periods, requiring updates to the existing recognition model using this new data. Therefore, the entire model update process needs to be divided into several incremental learning phases. Each stage independently acquires and processes newly added radar pulse sequences, thereby enabling continuous learning and capability expansion of the model.

[0053] Reference Figure 1 The implementation steps of this example include the following:

[0054] Step 1: Acquire data offline and divide it into training / validation data.

[0055] (1.1) In the incremental stage The receiver continuously intercepts pulse signals from multiple radars. These pulses interweave on the time axis, forming several aliased pulse streams. A total of [number missing] pulses were acquired during this phase. For each of these pulse sequences, parameter measurements are performed to obtain a pulse descriptor sequence:

[0056] ,

[0057] in, Indicates the first A descriptive word sequence composed of multiple radar pulses superimposed. For the first The description word of each pulse, express Number of pulses included This indicates the dimension of the measurement parameters contained in each pulse descriptor. In this example, five commonly used key parameters are used to describe the radar pulse: pulse arrival time (TOA), pulse width (PW), carrier frequency (RF), pulse amplitude (PA), and pulse angle of arrival (DOA). Each pulse is then represented as a five-dimensional vector descriptor. :

[0058] ;

[0059] (1.2) is the sequence of descriptors for each pulse. All pulses are labeled with their corresponding radar tags. This yields the newly acquired radar dataset for this stage. :

[0060] ,

[0061] in, For descriptor sequence A vector consisting of the category labels corresponding to all pulses in the vector. This indicates the first descriptor in the sequence. The category label corresponding to each pulse;

[0062] (1.3) Divide the labeled pulse descriptor sequences into training datasets according to a predetermined ratio. and validation dataset .

[0063] This example assumes, but is not limited to, a division ratio of 8:2.

[0064] Step 2: Split the training data into pure subsequences.

[0065] To support the subsequent generation of aliased playback samples based on clean pulse streams, this step first involves processing the training dataset... The pure pulse subsequences are separated by category, so that subsequent steps can use the pure pulse streams of each category as the basic units for constructing different aliasing combinations;

[0066] Then, on the training dataset Each aliased sequence in the data is based on the label sequence. Decomposition yields pure subsequences for each category. The extracted first Pure sequence for:

[0067] ,

[0068] in, Indicates matrix transpose. express Belongs to the category The number of pulses.

[0069] Step 3, manage the bank memory.

[0070] After obtaining the pure pulse subsequences for each category, these subsequences need to be added to a memory bank to achieve continuous accumulation across stages. Simultaneously, the occurrence probability of each category is updated based on the number of newly added subsequences in the current stage. This ensures that the memory bank reflects the historical distribution of each category and provides dynamic category sampling basis for the online synthesis of subsequent mixed samples. The implementation includes the following:

[0071] (3.1) For each observed category Using the set of existing historical subsequences With the newly extracted subsequences in the current stage Together they form the pure subsequence training pool for this category. :

[0072] ,

[0073] in, This indicates categories already existing in historical memory. Subsequence, its initial conditions , In order to be in Extracting categories The number of pure subsequences;

[0074] (3.2) The training pool of pure subsequences of all observed classes is used to form the stage. Pure subsequence training pool :

[0075] ,

[0076] in, For the stage The accumulated set of observed categories shows that the memory does not directly store fixed aliased samples, but uses pure pulse descriptor subsequences as basic units, thus providing flexible and efficient materials for subsequent online synthesis.

[0077] (3.3) Perform class probability update, that is, adjust the class probability according to the number of pure subsequences in each class. :

[0078] ,

[0079] in, Radar category The frequency of normalization in different historical stages Indicates the current stage number. Representation phase The number of training samples, Representation phase Category obtained The number of pure subsequences, This indicates the number of known categories at the current stage, i.e. The number of elements in the array.

[0080] Step 4: Generate replay samples based on category probabilities.

[0081] After completing the memory update and obtaining the class probabilities for the current stage. Next, using the pure subsequences of each category in the memory bank, the selection of categories and subsequences is controlled according to the category probability. By shifting, merging, and pruning the pulse arrival time, aliased playback samples for incremental training are generated. The implementation includes:

[0082] (4.1) Regarding the first The aliased playback samples appear in a distribution according to the category. from Extract a subset of categories ;

[0083] (4.2) For each selected category ,from Randomly select a pure subsequence , its origin The pulse description word is composed of 1 pulse, represented as:

[0084] ,

[0085] in For the first in the sequence The description word vector of each pulse;

[0086] (4.2) According to each pulse corresponding The attributes of the dimensional vector descriptor will be used to define the pure subsequence. In the parameter dimension, it can be regarded as being composed of It consists of a vector of parameters, which is represented as:

[0087] ,

[0088] Where the pulse arrival time vector for: ,

[0089] This indicates the pure subsequence The Middle The pulse arrival time of each pulse;

[0090] (4.3) to Perform a time shift to obtain the pulse arrival time sequence. :

[0091] ,

[0092] in, Indicates from 0 to Random perturbations are randomly selected within a certain range to obtain the corresponding pure subsequence with random time shift. :

[0093] ;

[0094] (4.4) All selected categories subsequence of Merge to obtain the first The sequence after merging the aliased playback samples :

[0095] ,

[0096] in, Indicates sequence merging;

[0097] (4.5) Merge the sequences All pulses are sorted according to their arrival time to obtain the sorted sequence. :

[0098] ,

[0099] in, This indicates that the pulses in the sequence are sorted according to their arrival time.

[0100] (4.6) For the sorted sequence The process involves pruning, retaining only the time intervals interleaved from multiple radar categories, to obtain the [number of] [sequences / time intervals]. A sample of aliased playback :

[0101] ,

[0102] in, This indicates that only time intervals interleaved from sequences of multiple radar categories are retained;

[0103] (4.7) Repeat steps (4.1) to (4.6) to generate the next aliasing playback sample until the number of aliasing playback samples reaches the predefined requirement for the number of aliasing playback samples, and obtain the aliasing playback dataset.

[0104] Step 5: Network training and validation.

[0105] Existing sorting and recognition neural networks include the dilated convolutional network DCN proposed by Chao Wang et al. in the paper "A Radar Signal Deinterleaving Method Based on Semantic Segmentation with Neural Network", the bidirectional recurrent neural network BRNN proposed by Mike Schuste et al. in the paper "Bidirectional Recurrent Neural Networks", the bidirectional long short-term memory network BLSTM proposed by A. Graves et al. in the paper "Framewise phoneme classification with bidirectional LSTM networks", and the multi-feature semantic analysis network MFSD proposed by Rouxuan Chen et al. in the paper "Radar Signal Deinterleaving With Multifeature Semantics and Modular Network Design".

[0106] This example uses, but is not limited to, MFSD as the neural network for the incremental training phase, and trains and verifies its performance based on aliased playback samples. Its implementation includes:

[0107] (5.1) Set the initial learning rate and the maximum number of training rounds. In this example, the initial learning rate is 0.001 and the maximum number of training rounds is 200.

[0108] (5.2) In each training round, aliased playback samples are input into the neural network in batches for iterative training:

[0109] (5.2.1) Randomly divide all aliased playback samples in the aliased playback dataset into multiple batches. In this example, each batch is divided into several batches. The method of dividing samples into batches;

[0110] (5.2.2) For the first Each batch of aliased playback samples, according to the preprocessing requirements of MFSD, randomly selects consecutive samples from each aliased playback sample. A fixed-length sequence of pulses Then, preprocessing is performed, that is, each fixed-length sequence is... Modify all arrival time parameters TOA to arrival time difference DTOA:

[0111] ,

[0112] in, and These represent the fixed-length sequence. The Middle Arrival time and arrival time difference of each pulse;

[0113] (5.2.3) The preprocessed first Each batch of aliased playback samples is input into the neural network, and the forward propagation of the neural network is performed to obtain the batch class confidence matrix output by the neural network. :

[0114] ,

[0115] in, Indicates the number of this batch The class confidence matrix of each aliased playback sample. Indicates the number of this batch In the aliased playback sample, the first The category confidence vector corresponding to each pulse. Indicates the number of this batch In the aliased playback sample, the first Each pulse belongs to category Confidence level;

[0116] (5.2.4) Calculate the loss based on the preset loss function. The loss function can be the cross-entropy loss function, the mean squared error loss function, or other loss functions that can be used for supervised learning. In this example, the cross-entropy loss function is used to calculate the loss. :

[0117] ,

[0118] in, Indicates the number of this batch In the aliased playback sample, the first Each pulse is a real label;

[0119] (5.2.5) Obtain the gradient through backpropagation;

[0120] (5.2.6) The network parameters are updated using a gradient-based approach. The optimization methods used to update the network parameters can include stochastic gradient descent, momentum gradient descent, RMSProp, Adam method, etc. This example uses the Adam optimization method.

[0121] (5.2.7) Repeat steps (5.2.1) to (5.2.6) until all batches of preprocessed aliased playback samples are input into the neural network to obtain the currently updated neural network;

[0122] (5.3) Validate the performance of the updated neural network using validation data to obtain the validation accuracy:

[0123] (5.3.1) For each validation sample in the validation dataset, according to the preprocessing requirements of MFSD, randomly select consecutive samples from each validation sample. The pulses constitute a fixed-length verification sequence, where the first pulse... A fixed-length verification sequence is denoted as Then, preprocessing is performed, that is, each fixed-length verification sequence... Modify all arrival time parameters TOA to arrival time difference DTOA:

[0124] ,

[0125] in, and These represent the fixed-length verification sequences. The Middle Arrival time and arrival time difference of each pulse;

[0126] (5.3.2) Each fixed-length verification sequence The data is sequentially input into the neural network, and the forward propagation of the neural network is performed to obtain the class confidence matrix output by the neural network. :

[0127] ,

[0128] in, Indicates the first The first of the validation samples The category confidence vector corresponding to each pulse. Indicates the first The first verification sample Each pulse belongs to category Confidence level;

[0129] (5.3.3) For each confidence vector Calculate the index of the maximum value to obtain the network prediction label corresponding to that pulse. :

[0130] ,

[0131] in, This represents the function for calculating the index of the maximum value;

[0132] (5.3.4) Calculate the verification accuracy OA:

[0133] ,

[0134] in, This indicates the number of samples in the validation dataset. Indicates the first The first verification sample The true label of a pulse This indicates an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise.

[0135] (5.4) Update the learning rate. During the training process, strategies such as fixed learning rate, segmented decay of learning rate or adaptive adjustment of learning rate based on verification performance can be adopted. In this embodiment, the adaptive learning rate adjustment method based on verification accuracy is adopted, that is, when the verification accuracy no longer improves after 10 consecutive times, the learning rate is decayed to 0.1 times the current learning rate.

[0136] (5.5) Repeat steps (5.2) to (5.4) until one of the following three conditions is met: the verification accuracy is higher than 99.9%, or the verification accuracy no longer improves after 30 consecutive verifications, or the number of training rounds reaches the maximum. Then the training ends and the trained incremental neural network is obtained.

[0137] Step 6: Real-time sorting and identification of label output.

[0138] After obtaining the trained incremental neural network, it is necessary to use this network to sort and recognize the real-time captured pulse descriptor sequences and output pulse-by-pulse classification labels. The implementation includes:

[0139] (6.1) The receiver intercepts pulse signals from multiple radars in real time and measures their parameters to obtain the pulse descriptor sequence to be sorted;

[0140] (6.2) According to the preprocessing requirements of the MFSD network, the pulse descriptor sequence to be sorted is divided into multiple pulse descriptor sequences of fixed length. Each fixed-length pulse descriptor sequence contains consecutive... Common methods for dividing multiple fixed-length pulse descriptor sequences include sliding window partitioning and non-overlapping partitioning. This example uses non-overlapping partitioning.

[0141] (6.3) Separate the fixed-length pulse descriptor sequence for each to be sorted Modify all arrival time parameters TOA to arrival time difference DTOA:

[0142] ;

[0143] (6.3) Input the modified fixed-length pulse descriptor sequence into the trained incremental neural network to obtain the class confidence matrix output by the neural network. :

[0144] ,

[0145] in, Indicates the first in the sequence The category confidence vector corresponding to each pulse. Indicates the first Each pulse belongs to category Confidence level;

[0146] (6.4) For each confidence vector Calculate the index of the maximum value to obtain the network prediction label corresponding to that pulse. :

[0147] ,

[0148] in, This represents the function for calculating the index of the maximum value;

[0149] (6.5) Repeat steps (6.1) to (6.4) until the real-time sorting and recognition task in this stage is completed;

[0150] Step 7, maintain the memory.

[0151] As the incremental phase progresses, the number of pure subsequences of each category in the memory increases continuously. If this is not controlled, it will lead to a gradual increase in sample storage costs and training time, affecting the efficiency and engineering usability of incremental sorting and recognition. Therefore, after extracting the deep features of pure subsequences, it is necessary to perform stratified sampling and selection of new and old pure subsequences based on representative scores. This will maintain the coverage of each category in the feature space and the effectiveness of sorting and recognition while controlling the size of the memory.

[0152] Reference Figure 2This step includes:

[0153] (7.1) Randomly select each pure subsequence of each category in the memory bank in turn. A series of consecutive pulses form a fixed-length purified subsequence. All arrival time parameters (TOA) in each fixed-length purified subsequence are modified to arrival time difference (DTOA). Then, all modified fixed-length purified subsequences are sequentially input into a trained incremental neural network. After the attention block in the MFSD network structure, deep features are directly output, including the category... The The deep features corresponding to a pure subsequence can be represented as:

[0154] ,

[0155] in, This represents the number of output channels of the attention block in the MFSD network structure. This indicates the number of pulses in the pure subsequence;

[0156] (7.2) Each deep feature The feature representation of each pure subsequence is obtained by averaging along the pulse dimension. :

[0157] ;

[0158] (7.3) Based on the current stage number Choose a reference sequence set:

[0159] If the current stage is the initial stage, then each category uses all the pure subsequences of that category as the reference sequence set;

[0160] like Then each category uses the pure subsequence of the previous stage of memory as the reference sequence set;

[0161] (7.4) Calculate the mean of the feature representations of all sequences in each class of reference sequence set to obtain the centroid of the reference feature for that class. :

[0162] ,

[0163] in, Indicates category The set of reference sequences, This indicates the calculation of the number of elements in the set;

[0164] (7.5) Calculate the Euclidean distance between the feature representation of each pure subsequence and the reference feature centroid of the class to which the subsequence belongs, and use it as the representative score of the pure subsequence. :

[0165] ,

[0166] in, This represents the L2 norm of a vector, which is the square root of the sum of the squares of the differences between the vector's dimensions.

[0167] (7.6) Divide the newly added pure subsequences of each category into multiple intervals according to their representative scores, and sample each interval separately, so that the proportion of the newly added pure subsequences of each category in the memory bank is satisfied. :

[0168] ,

[0169] in, Indicates category The number of stages that have occurred up to the current stage. To force the update ratio;

[0170] (7.7) Divide the historical pure subsequences of each category into multiple intervals according to their representative scores, and sample each interval separately, so that the proportion of the pure subsequences of that category in the memory bank that satisfy the condition that the historical pure subsequences are [percentage missing]. .

[0171] Example 2: Radar signal incremental sorting and identification device based on pulse stream synthesis and playback.

[0172] Reference Figure 3 This example includes: data acquisition module 1, pulse separation module 2, memory management module 3, playback sample generation module 4, network training module 5, online sorting and recognition module 6, and memory maintenance module 7. The playback sample generation module 4 includes: time shift submodule 41, sequence merging submodule 42, pulse sorting submodule 43, and sequence pruning submodule 44. The memory maintenance module 7 includes: deep feature extraction submodule 71, deep feature segmentation module 72, and hierarchical sampling module 73.

[0173] The working principle of the entire device is as follows:

[0174] The data acquisition module 1 is used to acquire the pulse descriptor sequence collected offline and perform label annotation and training / validation division, output training data to the pulse separation module 2, and output validation data to the network training module 5;

[0175] The pulse separation module 2 is used to separate the aliased pulses from the training data from the data acquisition module 1 according to the category label and generate pure subsequences of each category, and output the pure subsequences to the memory bank management module 3.

[0176] The memory bank management module 3 is used to add the pure subsequence from the pulse separation module 2 and the memory bank maintained by the memory bank maintenance module 7 in the previous stage to the memory bank and perform class probability update, and output the updated memory bank and class probability to the playback sample generation module 4.

[0177] The playback sample generation module 4 is used to extract pure sub-sequences from the memory bank according to the memory bank and class probabilities from the memory bank management module 3, and perform time shifting and interleaving to synthesize aliased playback samples. Specifically: the time shifting sub-module 41 is used to perform time shifting on the pure sub-sequences extracted from the memory bank according to the memory bank and class probabilities from the memory bank management module 3, and outputs randomly time-shifted pure sub-sequences to the sequence merging sub-module 42; the sequence merging sub-module 42 is used to merge multiple randomly time-shifted pure sub-sequences from the time shifting sub-module 41, and outputs the merged sequence to the pulse sorting sub-module 43; the pulse sorting sub-module 43 is used to pulse sort the merged sequence from the sequence merging sub-module 42 according to the pulse arrival time, and outputs the sorted sequence to the sequence trimming sub-module 44; the sequence trimming sub-module 44 is used to trim redundant pulses from the sorted sequence from the pulse sorting sub-module 43, retaining only the time intervals of interleaving between different class sequences, and outputs aliased playback samples to the network training module 5.

[0178] The network training module 5 is used to input the aliased playback samples from the playback sample generation module 4 into the neural network for iterative training, and to verify them using the verification data from the data acquisition module 1, and output the trained incremental neural network to the online sorting and recognition module 6 and the memory bank maintenance module 7.

[0179] The online sorting and identification module 6 is used to acquire test data in real time and input it into the trained incremental neural network from the network training module 5, and output pulse labels.

[0180] The memory bank maintenance module 7 is used to perform deep feature extraction, scoring, stratified sampling, and memory bank capacity update on all pure subsequences in the memory bank from the memory bank management module 3, based on the trained incremental neural network from the self-network training module 5. Specifically: the deep feature extraction submodule 71 is used to sequentially input all pure subsequences from the memory bank management module 3 into the trained incremental neural network from the self-network training module 5, obtain deep features, and output the deep features to the deep feature scoring module 72; the deep feature scoring module 72 is used to calculate representative scores based on the deep features from the deep feature extraction submodule 71 and output the representative scores to the stratified sampling module 73; the stratified sampling module 73 is used to perform stratified sampling and update the memory bank capacity based on the representative scores from the deep feature scoring module 72, and output the updated memory bank to the next stage of the memory bank management module 3.

[0181] It should be noted that the above functional modules can be implemented, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, they can be implemented, in whole or in part, as program instruction products. A program instruction product includes one or a set of program instructions. When the program instructions are loaded and executed on a computer, the described process or function is generated, in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The program instructions can be stored in a computer-readable and writable storage medium, or transferred from one computer's readable and writable storage medium to another.

[0182] The direct coupling or communication connections between the modules shown or discussed in this embodiment can be achieved through indirect coupling or communication connections via interfaces, devices, or modules. The various functional modules and sub-modules in this embodiment can dynamically reside within a single processing unit, or each module can exist physically independently, or two or more modules can dynamically reside within a single processing unit. When these dynamic components are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable and writable storage medium. This storage medium can be a memory, disk, or optical disc, etc.

[0183] The effects of this invention can be further illustrated by the following simulation results:

[0184] I. Simulation Conditions

[0185] The hardware environment is as follows: GPU (RTX 4090, 24 GB VRAM), CPU (64 cores), and 90 GB RAM.

[0186] The software environment is: Ubuntu 22.04, Python 3.10, PyTorch 2.1.2, CUDA 12.1.

[0187] The pulse descriptor parameter configurations are summarized in Table 1.

[0188] Table 1 Radar Parameter Settings

[0189]

[0190] Regarding specific parameter settings, unless otherwise specified, the range of PRI dwell modulation values ​​is as follows: The number of groups in a complete cycle is 4 to 6, and the number of pulses in each group is 4 to 6; the range of PRI fixed, staggered, and jitter modulation is all... The modulation parameter for the number of PRI values ​​within a single period in staggered modulation ranges from 3 to 10; PRI agility modulation is based on a predefined set of discrete values. ,exist The range contains 51 uniform step values ​​with a step size of [value missing]. Regarding PW, the PW range for all fixed modulations is set to... Regarding RF, the carrier frequency RF range of each radar is [missing information]. In carrier-frequency inter-pulse periodic agile modulation, the number of different RF values ​​within one cycle is 10 to 20. In carrier-frequency inter-pulse group periodic agile modulation, the number of pulses within each group is consistent with the number of pulses in each group during the pulse repetition interval, and the number of pulse groups within a complete cycle is 10 to 20. Regarding the amplitude PA, the PA range for non-scanning radar and phased-scanning radar is... The phased-scan radar has 4 beam positions in one complete cycle, and 4 PRI groups in each beam position; the mechanically scanned radar has a beamwidth factor of 11.5 and an amplitude reference... exist Inside. The DOA range of all radars is uniformly set to... All measurement parameters include measurement error.

[0191] In terms of incremental phase division, a total of 9 phases were set up to evaluate the incremental learning performance of the method. Among them, phase 1 is the initial training phase, and the remaining 8 phases are the continuous incremental process. Specifically, phases 3, 5, 7, and 9 introduce new models or patterns to simulate the continuous emergence of new radars or modulation patterns in the real environment; while phases 2, 4, 6, and 8 do not introduce new models or patterns, but some existing models or patterns will disappear from the training set, thus reflecting the situation where radars or patterns do not appear. In each incremental phase, a new model is a radar model that appears in this phase, a new pattern includes the pattern of the radar model that appears in this phase and the new appearance pattern of radar models that have appeared in previous phases, an old model is a radar model that has appeared in previous phases but has not appeared in this phase, an old pattern is an old model that has appeared in previous phases but has not appeared in this phase, a seen model includes the model seen in this phase and all models seen in previous phases, and a seen pattern includes the pattern seen in this phase and all patterns seen in previous phases. In all validation and test sets, all seen radars are assigned a fixed occurrence probability of 0.6, and one of their seen patterns is uniformly and randomly selected for generation. The training set for each incremental stage contains 10,000 samples, and the validation and test sets each contain 2,500 samples.

[0192] II. Simulation Content and Results

[0193] Simulation 1: Under the above simulation conditions, the incremental sorting and identification of radar signals was performed using the present invention and existing joint training methods, fine-tuning methods, GDumb methods, and LwF methods. The results were compared across various performance metrics. Figure 4 .in:

[0194] Figure 4 a) A graph showing the comparison of the average accuracy of all pulses;

[0195] Figure 4 b) A comparison chart of the cross-intersection and comparison indicators for all seen models;

[0196] Figure 4 c) A comparison chart of the average recall rate for all seen models;

[0197] Figure 4 d) A comparison chart of the average accuracy indicators for all models seen;

[0198] Figure 4 e) A graph showing the comparison results of the average recall rate of new models;

[0199] Figure 4 f) is a comparison chart of the average recall rate of older models;

[0200] Figure 4g) is a graph showing the comparison results of the average recall rate of the new styles;

[0201] Figure 4 h) is a graph showing the comparison results of the average recall rate of the old style;

[0202] from Figure 4 It is evident that the performance indicators of this invention in incremental sorting and identification of radar signals are superior to those of existing joint training methods, fine-tuning methods, GDumb methods, and LwF methods.

[0203] Simulation 2: Under the above simulation conditions, the incremental sorting and identification of radar signals was performed using the present invention and existing joint training methods, fine-tuning methods, GDumb methods, and LwF methods. The average time consumption of each training round was compared, and the results are as follows: Figure 5 .

[0204] from Figure 5 It is evident that the average training time per training round in the incremental sorting and identification of radar signals in this invention is significantly better than that of the joint training method. The training time does not increase significantly with the increase of the training stage, and the training time cost can be effectively controlled. Although it is similar to the fine-tuning method, GDumb method and LwF method, the performance index of this invention is significantly better than the fine-tuning method, GDumb method and LwF method, and the overall efficiency is high.

Claims

1. A radar signal incremental sorting and identification method based on pulse stream synthesis and playback, characterized in that, include (1) Obtain the radar signal pulse descriptor sequence collected offline, label it with radar tags, and divide it into training data and validation data; (2) Each pulse descriptor sequence in the training data is split into multiple pure subsequences according to the category label; (3) Add the pure subsequence to the class-based memory and perform class probability updates; (4) Extract pure subsequences of multiple categories from the memory bank according to the category probability, perform time shifting and interleaving synthesis to generate aliased playback samples; (5) Input the aliased playback samples into the neural network for iterative training, and verify them with the verification data to obtain the trained incremental neural network; (6) The radar signal pulse descriptor sequence to be sorted and identified is acquired in real time, and it is used as test data to be input into the trained incremental neural network. The output representation corresponding to each pulse is output, and the radar category corresponding to the maximum value is selected as the radar label of the pulse. (7) For a certain category of pure subsequences whose number in the memory bank exceeds the upper limit of the memory bank capacity, they are also input into the trained incremental neural network, output the deep features of the pure subsequences, score the deep features and perform stratified sampling, and update the memory bank capacity according to the stratified sampling results. (8) Repeat steps (1) to (7).

2. The method according to claim 1, characterized in that, The process of obtaining the radar signal pulse descriptor sequence acquired offline, labeling it with radar tags, and dividing it into training data and validation data includes: 1a) Obtain the pulse stream from the information intercepted by the receiver, and measure its parameters to obtain the pulse descriptor sequence; 1b) Label each pulse in each pulse descriptor sequence with its corresponding radar tag; 1c) Divide the labeled pulse descriptor sequences into training data and validation data according to a predetermined ratio.

3. The method according to claim 1, characterized in that, The class probability update performed in (3) is to adjust the class probability based on the number of pure subsequences in each class. : ， in, Radar category The frequency of normalization in different historical stages Indicates the current stage number. Representation phase The number of training samples, Representation phase Category obtained The number of pure subsequences, This indicates the number of known categories at the current stage.

4. The method according to claim 1, characterized in that, In step (4), the pure subsequences of multiple categories extracted from the memory are time-shifted and interleaved to generate aliased playback samples, which include: 4a) Time-shift the extracted pure subsequence to obtain the time-shifted pulse arrival time sequence. : , in, Indicates the type of radar extracted. The pulse arrival time series of the pure subsequence. and These represent the arrival times of the first and second pulses, respectively. This indicates the number of pulses in the pure subsequence. Indicates from 0 to Random disturbance quantities randomly selected within the range; 4b) First, merge the pure subsequences of multiple categories into one sequence. Then, sort all the pulses in the merged sequence according to the pulse arrival time. Finally, trim the sorted sequence to retain only the time intervals of the interleaved sequences from multiple radar categories to obtain the aliased playback sample.

5. The method according to claim 1, characterized in that, The step (5) involves inputting the aliased playback samples into the neural network for iterative training and validating them using validation data. This includes: 5a) Input the aliased playback samples into the neural network in batches, perform forward propagation for each batch in sequence, calculate the loss based on the preset loss function, obtain the gradient through back propagation, and update the network parameters in a gradient-based manner; 5b) Use validation data to validate the performance of the updated neural network and obtain the validation accuracy; 5c) Repeat steps 5a) and 5b) until the verification accuracy no longer improves, then end the training and obtain the trained incremental neural network.

6. The method according to claim 1, characterized in that, The process of scoring deep features and performing stratified sampling in (7), and updating the memory capacity based on the stratified sampling results, includes the following implementation: 7a) Input each pure subsequence in the memory bank into the trained incremental neural network to obtain deep features; 7b) Average each deep feature along the pulse dimension to obtain the feature representation of each pure subsequence; 7c) Select a set of reference sequences: If the current stage is the initial stage, then each category uses all the pure subsequences of that category as the reference sequence set; like Then each category uses the pure subsequence of the previous stage of memory as the reference sequence set; 7d) Calculate the mean of the feature representations of all sequences in each class of reference sequence set to obtain the reference feature centroid of that class; 7e) Calculate the Euclidean distance between the feature representation of each pure subsequence and the reference feature centroid of the class to which the subsequence belongs, and use it as the representative score of the pure subsequence; 7f) Divide the newly added pure subsequences of each category into multiple intervals according to their representative scores, and sample each interval separately, so that the proportion of newly added pure subsequences of that category in the memory bank is satisfied. : ， in, Indicates the current stage number. Indicates category The number of stages that have occurred up to the current stage. To force the update ratio; 7g) Divide the historical pure subsequences of each category into multiple intervals according to their representative scores, and sample from each interval to ensure that the proportion of historical pure subsequences in the memory bank for that category is [percentage missing]. .

7. A radar signal incremental sorting and identification device based on pulse stream synthesis and playback, characterized in that, include: The data acquisition module is used to acquire offline collected pulse descriptor sequences and perform label annotation and training / validation division; The pulse separation module is used to separate aliased pulses according to category labels and generate pure subsequences for each category. The memory management module is used to add pure subsequences to the memory and perform class probability updates; The playback sample generation module is used to extract pure subsequences from the memory bank based on class probability and perform time shifting and interleaving to synthesize aliased playback samples. The network training module is used to input aliased playback samples into the neural network for iterative training and to use validation data for validation, so as to obtain the incremental neural network after training. The online sorting and recognition module is used to input real-time test data into a trained incremental neural network and output pulse labels; The memory maintenance module is used for scoring based on deep features, hierarchical sampling, and updating the memory capacity.

8. The apparatus according to claim 7, characterized in that, The playback sample generation module includes: The time-shift submodule is used to time-shift the pure subsequences extracted from the memory. The sequence merging submodule is used to merge multiple time-shifted sequences; The pulse sorting submodule is used to sort all pulses according to their arrival time. The sequence trimming submodule is used to trim redundant pulses in a sequence, retaining only the time intervals where different types of sequences intersect.

9. The apparatus according to claim 7, characterized in that, The memory bank maintenance module includes: The deep feature extraction submodule is used to input the pure subsequence into the trained incremental neural network to obtain deep features; The deep feature scoring module is used to calculate representative scores based on deep features; The stratified sampling module is used to perform stratified sampling based on representative scores and update the memory capacity.

Images