A radar working mode open set identification method and device of a two-way autoencoder

Abstract

The application relates to a radar working mode open set recognition method and device of a two-way autoencoder, which comprises the following steps: inputting a to-be-recognized sample into a DA network trained to obtain a to-be-recognized feature vector; the DA network is obtained according to training data sets and a hybrid loss function; the training data sets are obtained according to original pulse sequences; the hybrid loss function is constructed according to a cross-entropy loss function, a center loss function and a reconstruction loss function; the distance between the to-be-recognized feature vector and a known category feature center is calculated to obtain a first distance; the known category feature center is obtained by inputting the training data sets into the DA network for training; and the recognition result of the to-be-recognized sample is obtained according to the first distance and a distance threshold. The DA network improves the distinguishability of different radar working modes, and the hybrid loss function makes the DA network have good generalization and robustness in a complex non-ideal scene.

Description

Technical Field

[0001] This invention relates to the field of radar signal processing, and more specifically, to a method and apparatus for open set identification of radar operating modes using a dual-channel self-encoder. Background Technology

[0002] Radar operating mode recognition is a crucial aspect of radar reconnaissance signal processing. It refers to the detailed analysis of the operating mode categories of individual radar emitters after acquiring sorted signals. With the development of radar technology, computer technology, and related hardware and software, multi-functional radar systems with higher dynamics and more complex signal waveforms have been widely applied and deployed. A multi-functional radar system is an advanced radar capable of rapid beam waveform agility and simultaneously performing multiple functions such as search, confirmation, tracking, identification, and guidance. During reconnaissance of non-cooperative multi-functional radars, their multi-mode nature makes it difficult to obtain a complete list of operating mode categories, often resulting in unknown modes. When the radar reconnaissance system lacks the ability to distinguish unknown modes, it may misclassify them as known modes, leading to information loss and even affecting radar countermeasures effectiveness. Therefore, it is essential to enable radar operating mode recognition methods to possess open-set learning capabilities.

[0003] From a reconnaissance perspective, radar operating modes can be defined as sequential pulse sequences with a fixed arrangement. Because radar operating modes can reflect, to some extent, the waveform parameter scheduling patterns of a radar system and the threat level of the radiation source, they have attracted considerable attention from researchers for many years. Generally, current radar operating mode recognition methods mainly fall into three categories: statistical analysis algorithms, primarily template matching and histogram methods; model inference-based algorithms, mainly using syntactic models, Hidden Markov Models, and predictive state representation models; and deep neural network algorithms, represented by networks such as Recurrent Neural Networks (RNNs), Temporal Convolutional Networks (TCNs), and Long Short-Term Memory (LSTM). Statistical analysis-based algorithms perform well when processing reconnaissance signals from early, relatively simple conventional radar systems, but their performance drops sharply when dealing with multi-functional radar signals with complex waveform parameters and higher dynamics. Model inference-based radar operating mode recognition methods can reveal the internal state patterns of multi-functional radar systems in a concise and insightful manner. However, when processing non-cooperative, multi-functional radar reconnaissance signals, there are problems such as difficulty in obtaining prior information and excessive model complexity. Radar operating mode recognition methods based on deep neural networks possess powerful capabilities for mining deep latent features, showing significant advantages over traditional methods when dealing with operating mode recognition problems in irrational situations containing missing or erroneous pulses. However, current deep learning-based radar operating mode recognition methods mainly operate on complete databases where all operating mode categories are available, with little research on open-set radar operating mode recognition problems in cases with incomplete databases. Summary of the Invention

[0004] To address the aforementioned problems in the prior art, this invention provides a radar operating mode open set identification method and apparatus using a dual-channel self-encoder.

[0005] According to a first aspect of the present invention, a radar operating mode open set identification method based on a dual-channel self-encoder is provided, the method comprising:

[0006] The sample to be identified is input into the trained DA network to obtain the feature vector to be identified of the sample; wherein, the DA network is trained based on the training dataset and the hybrid loss function; the training dataset is obtained based on the original pulse sequence; the hybrid loss function is constructed based on the cross-entropy loss function, the center loss function, and the reconstruction loss function;

[0007] The distance between the feature vector to be identified and the known category feature center is calculated to obtain the first distance; wherein, the known category feature center is obtained by training the DA network with the training dataset input.

[0008] The identification result of the sample to be identified is obtained based on the first distance and the distance threshold.

[0009] Optionally, before inputting the sample to be identified into the trained DA network to obtain the feature vector to be identified of the sample, the method further includes:

[0010] The training dataset is input into the DA network to train the DA network;

[0011] The DA network is iteratively optimized using the hybrid loss function until a preset convergence condition is met, resulting in the trained DA network and the known category feature centers.

[0012] Optionally, the hybrid loss function is expressed as follows:

[0013] Loss H =Loss ce +αLoss cl +β(Loss re_s +Loss re_p );

[0014] Among them, Loss H Represents the hybrid loss function, Loss ce Let Loss represent the cross-entropy loss function. cl Let Loss represent the central loss function. re_s Loss represents the semantic sequence reconstruction loss. re_p The original pulse sequence reconstruction loss is represented by α, where α represents the first hyperparameter and β represents the second hyperparameter.

[0015] Optionally, before inputting the training dataset into the DA network to train the DA network, the method further includes:

[0016] Semantic information is extracted from the preprocessed original pulse sequence to obtain a semantic sequence;

[0017] The original pulse sequence and the semantic sequence are normalized respectively to obtain the normalized original pulse sequence and the semantic sequence;

[0018] The normalized original pulse sequence and the semantic sequence are subjected to sliding window processing to obtain multiple sub-sequences;

[0019] The multiple subsequences are divided into the training dataset and the test dataset.

[0020] Optionally, before extracting semantic information from the preprocessed original pulse sequence to obtain a semantic sequence, the method further includes:

[0021] Perform histogram processing on the original pulse sequence to obtain the histogram corresponding to the original pulse sequence;

[0022] Obtain the PRI value that appears most frequently in the histogram;

[0023] The upper and lower bounds are obtained based on the PRI value;

[0024] The original pulse sequence is filtered according to the upper and lower bounds to obtain the preprocessed original pulse sequence.

[0025] Optionally, the semantic sequence includes the duty cycle, RF inter-pulse modulation features, and PRI modulation features corresponding to the original pulse sequence.

[0026] Optionally, the DA network includes a raw pulse sequence feature extraction subnetwork and a semantic sequence feature extraction subnetwork.

[0027] Optionally, the process of obtaining the distance threshold includes:

[0028] Calculate the second distance from each training sample in the training dataset to the known category feature center;

[0029] The distance threshold is obtained based on the box plot method and the second distance corresponding to each training sample.

[0030] Optionally, obtaining the identification result of the sample to be identified based on the first distance and the distance threshold includes:

[0031] The sample to be identified when the first distance is greater than the distance threshold is identified as an unknown working mode;

[0032] The sample to be identified if the first distance is less than or equal to the corresponding distance is identified as a known working mode;

[0033] The identification result of the sample to be identified is obtained based on the unknown working mode and the known working mode obtained from the identification.

[0034] According to a second aspect of the present invention, a radar operating mode open set identification device with a dual-channel self-encoder is provided, the device comprising:

[0035] The feature vector acquisition module is used to input the sample to be identified into the trained DA network to obtain the feature vector to be identified of the sample; wherein, the DA network is trained based on the training dataset and the hybrid loss function; the training dataset is obtained based on the original pulse sequence; the hybrid loss function is constructed based on the cross-entropy loss function, the center loss function, and the reconstruction loss function;

[0036] The distance acquisition module is used to calculate the distance between the feature vector to be identified and the known category feature center to obtain a first distance; wherein, the known category feature center is obtained by training the DA network by inputting the training dataset;

[0037] The identification module is used to obtain the identification result of the sample to be identified based on the first distance and the distance threshold.

[0038] The technical solution provided by this invention may include the following beneficial effects:

[0039] Through the above technical solution, the DA network achieves complementary feature information of different feature domains in the high-dimensional latent feature space, thereby improving the distinguishability of different radar operating modes. By constructing a hybrid loss function that includes center loss, cross-entropy loss and reconstruction loss, the DA network can learn features with high intra-class convergence and inter-class separability. This enables the network to have good generalization and robustness in complex and non-ideal scenarios with a high proportion of missing pulses and false pulses, and can identify radar operating modes even with incomplete databases.

[0040] Other features and advantages of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0041] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the following detailed description to explain the invention, but do not constitute a limitation thereof. In the drawings:

[0042] Figure 1 This is a flowchart illustrating an open set identification method for radar operating modes using a dual-channel self-encoder, according to an exemplary embodiment.

[0043] Figure 2 This is a schematic diagram of the structure of a DA network according to an exemplary embodiment.

[0044] Figure 3 This is a schematic diagram of the structure of a Crossformer network according to an exemplary embodiment.

[0045] Figure 4a This is a schematic diagram illustrating the result before pulse cleaning according to an exemplary embodiment.

[0046] Figure 4b This is a schematic diagram illustrating the result before pulse cleaning according to an exemplary embodiment.

[0047] Figure 5 This is a schematic diagram illustrating a PRI modulation feature according to an exemplary embodiment.

[0048] Figure 6 This is a schematic diagram of a loss curve according to an exemplary embodiment.

[0049] Figure 7 This is a schematic diagram illustrating an experimental result according to an exemplary embodiment.

[0050] Figure 8 This is a schematic diagram illustrating another experimental result according to an exemplary embodiment.

[0051] Figure 9 This is a schematic diagram illustrating another experimental result according to an exemplary embodiment.

[0052] Figure 10 This is a block diagram illustrating a radar operating mode open set identification device with a dual-channel self-encoder according to an exemplary embodiment. Detailed Implementation

[0053] Figure 1 This is a flowchart illustrating an open-set radar operating mode identification method using a dual-channel self-encoder according to an exemplary embodiment, such as... Figure 1 As shown, the method includes the following steps.

[0054] S101. Input the sample to be identified into the trained DA network to obtain the feature vector to be identified of the sample; wherein, the DA network is trained based on the training dataset and the hybrid loss function; the training dataset is obtained based on the original pulse sequence; the hybrid loss function is constructed based on the cross-entropy loss function, the center loss function and the reconstruction loss function.

[0055] Understandable, Figure 2 This is a schematic diagram of the structure of a DA network according to an exemplary embodiment, such as... Figure 2 As shown, the DA (Dual Autoencoder) network comprises a primitive pulse semantic feature extraction subnetwork and a semantic sequence feature extraction subnetwork. Both the primitive feature extraction subnetwork and the semantic feature extraction subnetwork are based on the Crossformer network. Figure 3 This is a schematic diagram illustrating the structure of a Crossformer network according to an exemplary embodiment, such as... Figure 3As shown, the Crossformer network contains an encoder and a decoder. The encoder contains a Dimension-Segment-Wise (DSW) encoding layer, a segment merging layer, and a two-stage attention layer. The decoder contains a Two-Stage Attention (TSA) layer, a series of normalization layers and multi-head attention layers, a series of normalization layers and a multi-layer perceptron module, and a fully connected layer.

[0056] The DSW encoding layer divides the input sequence into multiple segments of equal length along the time dimension to facilitate the extraction of dependencies across time and dimensions. The segment merging layer merges adjacent feature vectors (excluding those from the first layer) along the time dimension to obtain time series at different scales. The TSA layer is used to capture the time and dimension dependencies in multivariate time series. Therefore, the DA network consists of 2 DSW layers, 2 segment merging layers, 4 TSA layers, 2 multi-head attention and layer normalization layers, 2 normalization layers, a multilayer perceptron module, and 2 fully connected layers.

[0057] Optionally, prior to S101, the method may further include:

[0058] The training dataset is fed into the DA network to train the DA network;

[0059] The DA network is iteratively optimized using a hybrid loss function until the preset convergence condition is met, resulting in a trained DA network and known category feature centers.

[0060] Understandably, the specific training process of the DA network includes: inputting training samples from the training dataset sequentially into two DSW layers, two segment merging layers, four TSA layers, two multi-head attention layers, a layer normalization layer, two normalization layers, a multilayer perceptron module, and two fully connected layers; the two fully connected layers output deep latent features and deep semantic features respectively, and the output deep latent features and semantic features are fused by concat to obtain the final feature vector of each sample; the minimum Euclidean distance from the feature vector obtained in the previous step to the feature center vector of each known class is calculated, and the minimum Euclidean distance is combined into a distance vector, which is used as the input of the box plot method to calculate a distance threshold.

[0061] The minimum Euclidean distance from the feature vectors of the training samples obtained in the previous step to the center vectors of each known class feature is calculated using the following formula:

[0062]

[0063] in, Let N represent the known category feature centers, φ(·) represent the feature extractor, and N... kcLet P represent the number of known categories, c represent the working mode categories of the training samples, min(·) represent the function to find the minimum value, and dist(·) represent the function to calculate the Euclidean distance.

[0064] Alternatively, the hybrid loss function can be expressed as follows:

[0065] Loss H =Loss ce +αLoss cl +β(Loss re_s +Loss re_p );

[0066] Among them, Loss H Represents the mixed loss function, Loss ce Represents the cross-entropy loss function, Loss cl Represents the central loss function, Loss re_s Loss represents the semantic sequence reconstruction loss. re_p Let α represent the reconstruction loss of the original pulse sequence, β represent the first hyperparameter, and β represent the second hyperparameter.

[0067] Optionally, the process of obtaining the distance threshold includes:

[0068] Calculate the second distance from each training sample in the training dataset to the known category feature center;

[0069] The distance threshold is obtained based on the box plot method and the second distance corresponding to each training sample.

[0070] Understandably, the process of obtaining the distance threshold specifically includes: combining the minimum Euclidean distance (second distance) from each of the above training samples to the known class center into a distance vector, and arranging the distance vector in ascending order; calculating the upper quartile Q1 and lower quartile Q3 of the distance vector; calculating the interquartile range IQR = Q3 - Q1; and the distance threshold = Q3 + 1.5IQR.

[0071] Optionally, before feeding the training dataset into the DA network to train the DA network, the method further includes:

[0072] Semantic information is extracted from the preprocessed original pulse sequence to obtain a semantic sequence;

[0073] The original pulse sequence and semantic sequence are normalized respectively to obtain the normalized original pulse sequence and semantic sequence;

[0074] The normalized original pulse sequence and semantic sequence are subjected to sliding window processing to obtain multiple sub-sequences;

[0075] The multiple subsequences are divided into training and testing datasets.

[0076] It is understandable that the original pulse sequence is data from a single radar source obtained after sorting. Inevitably, there will be missing and incorrect pulses. Missing pulses disrupt the temporal relationship of the pulse sequence, while incorrect pulses simultaneously disrupt the temporal regularity and distribution characteristics of the original pulse sequence. Therefore, preprocessing of the original pulse sequence is necessary.

[0077] Furthermore, to address the problem that existing radar operating mode recognition methods struggle to effectively identify unknown operating modes under incomplete sample database conditions, the constructed DA network can input the original pulse sequence and semantic sequence into the DA network to extract corresponding deep latent features. The deep latent features extracted from the original pulse sequence and the deep semantic features extracted from the semantic sequence are then fused, which facilitates the complementarity of different features and improves the separability of different operating mode categories in the high-dimensional feature space.

[0078] The Pulse Repetition Interval (PRI) is defined as the time difference between the arrival of adjacent pulses and has good stability. Figure 4a This is a schematic diagram illustrating the result before pulse cleaning according to an exemplary embodiment. Figure 4b This is a schematic diagram illustrating the result before pulse cleaning according to an exemplary embodiment.

[0079] Optionally, the semantic sequence includes the duty cycle, RF inter-pulse modulation features, and PRI modulation features corresponding to the original pulse sequence.

[0080] It is understandable that duty cycle, RF (radio frequency) inter-pulse modulation features, and PRI modulation features are all semantic information tightly coupled with radar operating modes. By extracting this semantic information, prior radar knowledge can be introduced into the open-set recognition task of radar operating mode identification, which is of great significance for improving recognition accuracy.

[0081] Calculate the duty cycle sequence using the following formula:

[0082]

[0083] Where L represents the total number of pulses, i represents the pulse sequence number, and d i PW represents the duty cycle of the i-th pulse. i PRI represents the pulse width of the i-th pulse. i This represents the PRI value of the i-th pulse.

[0084] RF and PRI are crucial parameters tightly coupled with radar operating modes. For example, radar systems often employ RF-agile inter-pulse modulation types to enhance radar anti-jamming capabilities, and use PRI-variable inter-pulse modulation types to resolve range or velocity ambiguity. Therefore, the choice of RF and PRI inter-pulse modulation types is closely related to the radar operating mode. To effectively distinguish the differences between different inter-pulse modulation types, these types can be converted into specific feature vectors for characterization. Introducing the sequence {κ... i}, its i-th element can be represented as:

[0085]

[0086] Where, △x i Let x represent the first difference of the i-th element. i RF i or PRI i ε′ represents the corresponding threshold, ε′=0.01σ x , σ x This represents the mean of the corresponding sequence. Furthermore, we can define an eigenvector V. i as follows:

[0087]

[0088] The eigenvectors V of RF and PRI can be obtained using the above process. rf and V pri .

[0089] Figure 5 This is a schematic diagram illustrating a PRI modulation feature according to an exemplary embodiment, such as... Figure 5 As shown, the PRI modulation characteristics corresponding to different PRI modulation types are illustrated. It can be seen that the differences in characteristics among different PRI inter-pulse modulation types are significant, which is of great value for understanding subsequent radar operating modes.

[0090] In one implementation, since the dimensions of parameters for different radar operating modes are different, in order to avoid the influence of dimensions and improve the training and convergence speed of the network, each parameter needs to be normalized to the range [0,1] according to the following formula.

[0091]

[0092] Where x represents the parameter vector that needs to be normalized, and x′ represents the normalized vector.

[0093] When performing sliding window processing on the normalized original pulse sequence and semantic sequence, a non-overlapping sliding window can be used to segment the original radar working mode sequence and semantic information sequence according to a preset length, resulting in multiple sub-sequences.

[0094] A portion of the existing working pattern categories are designated as known categories, and the remaining portion as unknown categories. The number of known and unknown categories can be preset. Multiple subsequences are then divided into training and testing datasets.

[0095] Optionally, before extracting semantic information from the preprocessed original pulse sequence to obtain a semantic sequence, the method further includes:

[0096] Histogram processing is performed on the original pulse sequence to obtain the histogram corresponding to the original pulse sequence;

[0097] Obtain the PRI value that appears most frequently in the histogram;

[0098] The upper and lower bounds can be determined based on the PRI value;

[0099] The original pulse sequence is filtered based on the upper and lower bounds to obtain the preprocessed original pulse sequence.

[0100] Understandably, the original pulse sequence is preprocessed by cleaning based on the PRI parameters. The specific steps are as follows: The PRI sequence of the original pulse sequence is processed using a histogram, with a statistical interval of 1µs; the PRI value with the highest frequency in the histogram is found, and 10% of this value is taken as the lower bound, and 200% as the upper bound; the original PRI pulse sequence is filtered according to the upper and lower bounds of the pulse PRI values. When the values ​​are within the lower and upper bounds, the parameters of the original pulse are retained; if the values ​​are greater than the upper bound or less than the lower bound, the values ​​of the corresponding pulse parameters are set to 0. Pulse cleaning can, to some extent, suppress the influence of missing or erroneous pulses on the radar's operating mode patterns.

[0101] S102. Calculate the distance between the feature vector to be identified and the known category feature center to obtain the first distance; wherein, the known category feature center is obtained by training the DA network by inputting the training dataset.

[0102] S103. Obtain the recognition result of the sample to be identified based on the first distance and the distance threshold.

[0103] Optionally, S104 includes:

[0104] The sample to be identified that has a first distance greater than the distance threshold is identified as an unknown working mode;

[0105] The sample to be identified with a first distance less than or equal to the corresponding working mode is identified as a known working mode;

[0106] The identification results of the sample to be identified are obtained based on the unknown working modes and known working modes obtained from the identification.

[0107] Understandably, the first distance to the sample to be identified is compared with the distance threshold, and the corresponding classifier result is used as the final identification result.

[0108] In one implementation, the simulation conditions were as follows: the experiment was conducted under different levels of missing and false pulses, simulating data from 13 operating modes of the airborne multi-function radar, including TWS, TAS, VS, RWS, Multiple Target Track (MTT), Single Target Track (STT), Situation Awareness Mode (SAM), High Power Jamming (HPJ), Velocity Search Range (VSR), Weather Avoidance (WA), Sea Surface Target Tracking (SSTT), Sea Surface Search (SSS), and Target Identification (TID). The parameter ranges for different operating modes are shown in Table 1. Each operating mode generated 120,000 pulses; the ratio of missing and false pulses was set between 10% and 40%; measurement error was achieved by adding Gaussian white noise with a mean of 0 and a variance of 2 μs to each pulse.

[0109] Table 1

[0110] Work mode RF (GHz) PW(μs) PRI(μs) Duty cycle (%) TWS 9.5～9.9 2～60 50～333 0.1～25 TAS 9.5～9.9 2～60 50～200 0.1～25 VS 9.5～9.9 1～3 3.33～10 10～30 STT 9.5～9.9 0.1～10 50～500 0.1～10 RWS 9.5～9.9 1～20 50～166.67 1～25 VSR 9.5～9.9 1～20 50～166.67 1～25 MTT 9.5～9.9 0.1～20 3.33～125 0.1～25 HPJ 9.5～9.9 3～10 3.33～20 10～50 WA 9.5～9.9 1～50 200～2000 1～10 SSS 9.5～9.9 1～200 500～2000 0.1～10 SSTT 9.5～9.9 1～200 50～500 0.1～10 TID 9.5～9.9 1～200 50～500 0.1～10 SAM 9.5～9.9 0.1～10 50～500 0.1～10

[0111] The experimental setup is as follows: The parameters of the frequency repetition type recognition network used in this invention are shown in Table 2. In Table 2, Batch size is the number of samples used to update the model weights in each iteration, epoch is the iteration batch, Learningrate is the learning rate, optimizer is the optimizer, and Weight decay is the weight decay coefficient, a parameter to prevent overfitting. 80% of all samples are randomly selected for training, and the remaining 20% ​​are used for validation and testing.

[0112] Table 2

[0113]

[0114]

[0115] Figure 6 This is a schematic diagram of a loss curve according to an exemplary embodiment. As can be seen, the network achieved good convergence after 300 iterations.

[0116] Figure 7 This is a schematic diagram illustrating an experimental result according to an exemplary embodiment, such as... Figure 7 As shown, Figure 7 These are the experimental results under the condition of only missing pulses. It can be seen that as the number of missing pulses gradually increases, the recognition result decreases to some extent. This is because the presence of missing pulses disrupts the parameter regularity of different working mode sequences, leading to overlap of extracted deep features in high-dimensional space. However, the overall recognition result in all cases is better than 83%, indicating that the DA algorithm is not sensitive to missing pulses.

[0117] Figure 8 This is a schematic diagram illustrating another experimental result according to an exemplary embodiment, such as... Figure 8 As shown, Figure 8 The experimental results are presented for the case with only erroneous pulses. Overall, the recognition results are lower in the case of erroneous pulses compared to the case with only missing pulses. This is because when the pulse sequence contains erroneous pulses, the temporal regularity of different working mode sequences is disrupted, and new error attribute parameter information is introduced, making the situation more complex. However, the experimental results effectively demonstrate that the DA algorithm has good generalization ability for the case of erroneous pulses.

[0118] Figure 9 This is a schematic diagram illustrating another experimental result according to an exemplary embodiment, such as... Figure 9 As shown, Figure 9 The experimental results, including both missing and erroneous pulses, show a significant drop in overall performance when the proportion of both exceeds 20%. This is because the simultaneous presence of missing and erroneous pulses severely disrupts the parameter patterns of different operating modes. However, due to the use of the DA network, feature information from different dimensions can be fused to achieve a certain degree of complementarity, thus ensuring stable recognition performance. Therefore, even when the proportion of both missing and erroneous pulses reaches as high as 40%, the overall recognition result still remains around 80%. These results demonstrate that the DA network exhibits good generalization ability in complex environments.

[0119] Figure 10 This is a block diagram illustrating a radar operating mode open set identification device with a dual-channel self-encoder according to an exemplary embodiment, such as... Figure 10 As shown, the radar operating mode open set identification device 1000 of the dual-channel self-encoder may include:

[0120] The feature vector acquisition module 1001 is used to input the sample to be identified into the trained DA network to obtain the feature vector to be identified of the sample; wherein, the DA network is trained based on the training dataset and the hybrid loss function; the training dataset is obtained based on the original pulse sequence; the hybrid loss function is constructed based on the cross-entropy loss function, the center loss function and the reconstruction loss function;

[0121] The distance acquisition module 1002 is used to calculate the distance between the feature vector to be identified and the known category feature center to obtain the first distance; wherein, the known category feature center is obtained by training the DA network by inputting the training dataset;

[0122] The recognition module 1003 is used to obtain the recognition result of the sample to be recognized based on the first distance and the distance threshold.

[0123] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0124] The preferred embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the specific details of the above embodiments. Within the scope of the technical concept of the present invention, various simple modifications can be made to the technical solution of the present invention, and these simple modifications all fall within the protection scope of the present invention.

[0125] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, the present invention will not describe the various possible combinations separately.

[0126] Furthermore, various different embodiments of the present invention can be combined in any way, as long as they do not violate the spirit of the present invention, they should also be regarded as the content disclosed by the present invention.

Claims

1. A radar operating mode open set identification method using a dual-channel self-encoder, characterized in that, The method includes: The pulse sequence to be identified is input into the trained DA network to obtain the feature vector to be identified for the pulse sequence. The DA network is trained using a training dataset and a hybrid loss function. The training dataset is obtained from the original pulse sequence. The hybrid loss function is constructed using cross-entropy loss, center loss, and reconstruction loss functions. The pulse sequence to be identified is obtained from data of various operating modes of an airborne multi-function radar. The DA network is a dual-channel autoencoder network. The DA network includes a sub-network for extracting semantic features from the original pulse and a sub-network for extracting semantic sequence features. Both the original feature extraction sub-network and the semantic feature extraction sub-network use CrossForward algorithm. The Crossformer network is based on the Mer network. It comprises an encoder and a decoder. The encoder includes a dimensional segmented encoding layer, a segment merging layer, and a two-stage attention layer. The decoder includes a two-stage attention layer, a layer normalization layer, a multi-head attention layer, a layer normalization layer, a multilayer perceptron module, and a fully connected layer. The dimensional segmented encoding layer divides the input sequence into multiple segments of equal length according to the time dimension. The segment merging layer merges adjacent feature vectors (excluding the first layer) according to the time dimension. The two-stage attention layer captures the temporal and dimensional dependencies in the multivariate time series. The distance between the feature vector to be identified and the known category feature center is calculated to obtain the first distance; wherein, the known category feature center is obtained by training the DA network with the training dataset input. The identification result of the pulse sequence to be identified is obtained based on the first distance and the distance threshold; The process of obtaining the distance threshold includes: Calculate the second distance from each training sample in the training dataset to the known category feature center; The distance threshold is obtained based on the box plot method and the second distance corresponding to each training sample; The step of obtaining the identification result of the pulse sequence to be identified based on the first distance and the distance threshold includes: The pulse sequence to be identified corresponding to a first distance greater than the distance threshold is identified as an unknown working mode; The first distance is less than or equal to the corresponding pulse sequence to be identified as a known working mode; The identification result of the pulse sequence to be identified is obtained based on the unknown working mode and the known working mode obtained from the identification.

2. The radar operating mode open set identification method of dual-channel self-encoder according to claim 1, characterized in that, Before inputting the pulse sequence to be identified into the trained DA network to obtain the feature vector to be identified of the pulse sequence, the method further includes: The training dataset is input into the DA network to train the DA network; The DA network is iteratively optimized using the hybrid loss function until a preset convergence condition is met, resulting in the trained DA network and the known category feature centers.

3. The radar operating mode open set identification method of dual-channel self-encoder according to claim 2, characterized in that, The hybrid loss function is expressed as follows: ； in, This represents the hybrid loss function. This represents the cross-entropy loss function. Denotes the central loss function, Represents the semantic sequence reconstruction loss. This represents the reconstruction loss of the original pulse sequence. Indicates the first hyperparameter. This represents the second hyperparameter.

4. The radar operating mode open set identification method of the dual-channel self-encoder according to claim 2, characterized in that, Before inputting the training dataset into the DA network to train the DA network, the method further includes: Semantic information is extracted from the preprocessed original pulse sequence to obtain a semantic sequence; The original pulse sequence and the semantic sequence are normalized respectively to obtain the normalized original pulse sequence and the semantic sequence; The normalized original pulse sequence and the semantic sequence are subjected to sliding window processing to obtain multiple sub-sequences; The multiple subsequences are divided into the training dataset and the test dataset.

5. The radar operating mode open set identification method of dual-channel self-encoder according to claim 4, characterized in that, Before extracting semantic information from the preprocessed original pulse sequence to obtain a semantic sequence, the method further includes: Perform histogram processing on the original pulse sequence to obtain the histogram corresponding to the original pulse sequence; Obtain the PRI value that appears most frequently in the histogram; The upper and lower bounds are obtained based on the PRI value; The original pulse sequence is filtered according to the upper and lower bounds to obtain the preprocessed original pulse sequence.

6. The radar operating mode open set identification method of dual-channel self-encoder according to claim 4, characterized in that, The semantic sequence includes the duty cycle, RF inter-pulse modulation features, and PRI modulation features corresponding to the original pulse sequence.

7. The radar operating mode open set identification method of dual-channel self-encoder according to claim 1, characterized in that, The DA network includes a sub-network for extracting features from the original pulse sequence and a sub-network for extracting features from the semantic sequence.

8. A radar operating mode open set identification device with dual-channel self-encoder, characterized in that, The device includes: The feature vector acquisition module is used to input the pulse sequence to be identified into the trained DA network to obtain the feature vector to be identified of the pulse sequence; wherein, the DA network is trained based on a training dataset and a hybrid loss function; the training dataset is obtained based on the original pulse sequence; the hybrid loss function is constructed based on the cross-entropy loss function, the center loss function, and the reconstruction loss function; the pulse sequence to be identified is obtained based on the operating mode data of the airborne multi-function radar; the DA network is a dual-channel autoencoder network; the DA network includes an original pulse semantic feature extraction subnetwork and a semantic sequence feature extraction subnetwork; both the original feature extraction subnetwork and the semantic feature extraction subnetwork use Cr The Crossformer network is based on a multivariate time series model. The Crossformer network includes an encoder and a decoder. The encoder includes a dimensional segmentation encoding layer, a segment merging layer, and a two-stage attention layer. The decoder includes a two-stage attention layer, a layer normalization layer, a multi-head attention layer, a layer normalization layer, a multilayer perceptron module, and a fully connected layer. The dimensional segmentation encoding layer divides the input sequence into multiple segments of equal length according to the time dimension. The segment merging layer merges adjacent feature vectors (excluding those in the first layer) according to the time dimension. The two-stage attention layer captures the temporal and dimensional dependencies in the multivariate time series model. The distance acquisition module is used to calculate the distance between the feature vector to be identified and the known category feature center to obtain a first distance; wherein, the known category feature center is obtained by training the DA network by inputting the training dataset; The identification module is used to obtain the identification result of the pulse sequence to be identified based on the first distance and the distance threshold; The process of obtaining the distance threshold includes: Calculate the second distance from each training sample in the training dataset to the known category feature center; The distance threshold is obtained based on the box plot method and the second distance corresponding to each training sample; The identification module is further configured to identify the pulse sequence to be identified corresponding to the first distance being greater than the distance threshold as an unknown working mode; The first distance is less than or equal to the corresponding pulse sequence to be identified as a known working mode; The identification result of the pulse sequence to be identified is obtained based on the unknown working mode and the known working mode obtained from the identification.

Images