Radar working mode open set identification method and device of double-path auto-encoder

Abstract

The invention relates to a radar working mode open set identification method and device for a double-path auto-encoder, and the method comprises the steps: inputting a to-be-identified sample into a trained DA network, and obtaining a to-be-identified feature vector; wherein the DA network is obtained by training according to a training data set and a mixed loss function; the training data set is obtained according to an original pulse sequence; the mixed loss function is constructed according to a cross entropy loss function, a center loss function and a reconstruction loss function; calculating the distance between the to-be-identified feature vector and the known category feature center to obtain a first distance; wherein the known category feature center is obtained by inputting a training data set into a DA network for training; and obtaining an identification result of the to-be-identified sample according to the first distance and a distance threshold. According to the method, the distinguishability of different radar working modes is improved through the DA network, and the DA network has good generalization and robustness in a complex non-ideal scene through the mixed loss function.

Description

Technical Field

[0001] The present invention relates to the field of radar signal processing, and in particular to an open-set recognition method and device for a radar operating mode of a dual-path autoencoder. Background Art

[0002] Radar working mode recognition is an important part of radar reconnaissance signal processing. It refers to the process of analyzing the working mode category in detail based on the signal of a single radar radiation source after sorting. With the development of radar technology, computer technology and related software and hardware technology, multi-function radar systems with higher dynamics and more complex signal waveforms have been widely used and deployed. Multi-function radar system is an advanced radar that can perform rapid beam waveform agility and can simultaneously perform multiple functions such as search, confirmation, tracking, identification, and guidance. In the process of reconnaissance of non-cooperative multi-function radars, due to their multi-mode characteristics, it is difficult to obtain their complete working mode categories, and they often face unknown modes. When the radar reconnaissance system does not have the ability to distinguish unknown modes, it will misclassify the unknown mode as a known mode, resulting in information loss and even affecting the radar countermeasure effect. Therefore, it is crucial to enable the radar working mode recognition method to have open set learning capabilities.

[0003] From the perspective of reconnaissance, the radar working mode can be defined as a sequential pulse sequence with a fixed arrangement order. Since the radar working mode can reflect the radar system waveform parameter scheduling law and the threat level of the radiation source to a certain extent, it has attracted much attention from researchers for many years. In general, the current radar working mode recognition methods mainly include three categories: statistical analysis algorithms with template matching method and histogram method as the mainstream; model inference-based algorithms with syntactic model, hidden Markov model and predictive state representation model as the main ones; deep neural network algorithms represented by networks such as recurrent neural network (RNN), temporal convolutional network (TCN), long short-term memory network (LSTM). The statistical analysis-based algorithms are effective in processing conventional radar reconnaissance signals with relatively simple parameters and functions in the early stage, but the performance drops sharply for multi-function radar signals with complex waveform parameters and relatively higher dynamics. The radar working mode recognition method based on model inference can deeply reveal the internal state law of the multi-function radar system in a concise way. However, when dealing with non-cooperative multi-function radar reconnaissance signals, there are problems such as difficulty in obtaining prior information and high model complexity. The radar working mode recognition method based on deep neural network has a powerful ability to mine deep implicit features, and has significant advantages over traditional methods when dealing with working mode recognition problems in irrational situations including missed and wrong pulses. However, the current radar working mode recognition method based on deep learning is mainly carried out on a complete database in which all working mode categories can be obtained, and there is little research on the open set recognition problem of radar working modes in the case of incomplete databases. Summary of the invention

[0004] In order to solve the above problems existing in the prior art, the present invention provides a radar operating mode open set identification method and device of a dual-path autoencoder.

[0005] According to a first aspect of an embodiment of the present invention, a method for open set identification of radar operating modes of a dual-path autoencoder is provided, the method comprising:

[0006] Inputting the sample to be identified into the trained DA network to obtain the feature vector to be identified of the sample to be identified; wherein the DA network is trained according to the training data set and the mixed loss function; the training data set is obtained according to the original pulse sequence; the mixed loss function is constructed according to the cross entropy loss function, the center loss function and the reconstruction loss function;

[0007] Calculating the distance between the feature vector to be identified and the feature center of a known category to obtain a first distance; wherein the feature center of the known category is obtained by inputting the training data set into the DA network for training;

[0008] The recognition result of the sample to be recognized is obtained according to 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 to be identified, the method further includes:

[0010] Inputting the training data set 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 reached, thereby obtaining the trained DA network and the known category feature center.

[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 mixed loss function, Loss ce Represents the cross entropy loss function, Loss cl Represents the central loss function, Loss re_s Represents the semantic sequence reconstruction loss, Loss re_p represents the original pulse sequence reconstruction loss, α represents the first hyperparameter, and β represents the second hyperparameter.

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

[0016] Extracting semantic information from the preprocessed original pulse sequence to obtain a semantic sequence;

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

[0018] Performing sliding window processing on the normalized original pulse sequence and the semantic sequence to obtain multiple subsequences;

[0019] The plurality of subsequences are divided into the training data set and the test data set.

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

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

[0022] Obtaining the PRI value with the highest frequency in the histogram;

[0023] Obtain upper and lower bounds based on the PRI value;

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

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

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

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

[0028] Calculate and obtain a second distance from each training sample in the training data set to the center of the known category feature;

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

[0030] Optionally, obtaining the recognition result of the sample to be recognized according to the first distance and the distance threshold includes:

[0031] Identify the sample to be identified corresponding to the first distance being greater than the distance threshold as an unknown working mode;

[0032] Identify the first distance being less than or equal to the corresponding to-be-identified sample as a known working mode;

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

[0034] According to a second aspect of an embodiment of the present invention, there is provided a radar operating mode open set identification device of a dual-path autoencoder, the device comprising:

[0035] A 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 to be identified; wherein the DA network is trained according to the training data set and the mixed loss function; the training data set is obtained according to the original pulse sequence; the mixed loss function is constructed according to the cross entropy loss function, the center loss function and the reconstruction loss function;

[0036] A distance acquisition module is used to calculate the distance between the feature vector to be identified and the feature center of a known category to obtain a first distance; wherein the feature center of the known category is obtained by inputting the training data set into the DA network for training;

[0037] An identification module is used to obtain an identification result of the sample to be identified according to the first distance and the distance threshold.

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

[0039] Through the above technical scheme, the DA network is used to realize the complementarity of feature information in different feature domains in the high-dimensional latent feature space, thereby improving the distinguishability of different radar working modes. By constructing a mixed loss function including center loss, cross entropy loss and reconstruction loss, the DA network can learn features with high intra-class convergence and inter-class separability, so that the network has good generalization and robustness in complex non-ideal scenarios with a high proportion of missed pulses and false pulses, and can identify radar working modes in the case of an incomplete database.

[0040] Other features and advantages of the present invention will be described in detail in the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the following specific embodiments, they are used to explain the present invention but do not constitute a limitation of the present invention. In the accompanying drawings:

[0042] Figure 1 The present invention is a flowchart of an open-set identification method of a radar operating mode of a dual-channel autoencoder according to an exemplary embodiment.

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

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

[0045] Figure 4a FIG. 1 is a schematic diagram showing a result before pulse cleaning according to an exemplary embodiment.

[0046] Figure 4b FIG. 1 is a schematic diagram showing a result before pulse cleaning according to an exemplary embodiment.

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

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

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

[0050] Figure 8 is another schematic diagram of experimental results according to an exemplary embodiment.

[0051] Fig. 9 is another schematic diagram of experimental results according to an exemplary embodiment.

[0052] Fig.10 It is a block diagram of an open-set recognition device for radar operating modes of a dual-path autoencoder according to an exemplary embodiment. DETAILED DESCRIPTION

[0053] Figure 1 is a flow chart of an open set identification method for radar working mode of a dual-path autoencoder according to an exemplary embodiment. 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 to be identified; wherein the DA network is trained according to the training data set and the mixed loss function; the training data set is obtained according to the original pulse sequence; the mixed loss function is constructed according to the cross entropy loss function, the center loss function and the reconstruction loss function.

[0055] Understandably, Figure 2 is a schematic diagram of a DA network structure according to an exemplary embodiment. Figure 2 As shown in FIG. 1 , the DA (Dual Autoencoder) 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 are based on the Crossformer network. Figure 3 is a schematic diagram of a structure of a Crossformer network according to an exemplary embodiment, such as Figure 3As shown in the figure, 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), layers of normalization layers and multi-head attention layers, layers of normalization layers and multi-layer perceptron modules, and a fully connected layer.

[0056] Among them, the DSW encoding layer divides the input sequence into multiple sub-segments of the same length according to the time dimension to facilitate the extraction of cross-time and cross-dimensional dependencies. The segment merging layer merges two adjacent feature vectors except the first layer according to the time dimension to obtain time series of different scales. The TSA layer is used to capture the time and dimension dependencies in multivariate time series. Therefore, the DA network includes 2 DSW layers, 2 segment merging layers, 4 TSA layers, 2 multi-head attention and layer normalization layers, 2 normalization layers and multi-layer perceptron modules, and 2 fully connected layers.

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

[0058] Input the training data set 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 reached, and the trained DA network and the known category feature centers are obtained.

[0060] It can be understood that the specific training process of the DA network includes: inputting the training samples in the training data set into 2 DSW layers, 2 segment merging layers, 4 TSA layers, 2 multi-head attention and layer normalization layers, 2 normalization layers and multi-layer perceptron modules and 2 fully connected layers in sequence; the 2 fully connected layers output deep latent features and deep semantic features respectively, and the output deep latent features and semantic features are concat fused to obtain the final feature vector of each sample; respectively calculate the minimum Euclidean distance from the feature vector obtained in the previous step to the feature center vector of each known class, and at the same time, combine the minimum Euclidean distance into a distance vector, and use it as the input of the box plot method to calculate a distance threshold.

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

[0062]

[0063] in, represents the known category feature center, φ(·) represents the feature extractor, N kcrepresents the number of known categories, P represents training samples, c represents the working mode category of training samples, min(·) represents the function of finding the minimum value, and dist(·) represents the function of calculating the Euclidean distance.

[0064] Optionally, the hybrid loss function is 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 Represents the semantic sequence reconstruction loss, Loss re_p represents the original pulse train reconstruction loss, α represents the first hyperparameter, and β represents the second hyperparameter.

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

[0068] Calculate the second distance between each training sample in the training data set and the center of the known category feature;

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

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

[0071] Optionally, before inputting the training data set into the DA network to train the DA network, the method further comprises:

[0072] Extracting semantic information from the preprocessed original pulse sequence to obtain a semantic sequence;

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

[0074] Perform sliding window processing on the normalized original pulse sequence and semantic sequence to obtain multiple subsequences;

[0075] Divide multiple subsequences into training and testing datasets.

[0076] It can be understood that the original pulse sequence is the data of a single radar radiation source obtained after sorting. Among them, there will inevitably be missed pulses and wrong pulses. The existence of missed pulses will destroy the timing relationship of the pulse sequence, and the existence of wrong pulses will simultaneously destroy the timing law and distribution characteristics of the original pulse sequence. Therefore, it is necessary to preprocess the original pulse sequence.

[0077] In addition, in response to the problem that existing radar working mode recognition methods are difficult to effectively identify unknown working modes under incomplete sample libraries, the constructed DA network can input the original pulse sequence and semantic sequence into the DA network to extract the corresponding deep implicit features, and fuse the deep implicit features extracted from the original pulse sequence and the deep semantic features extracted from the semantic sequence, which is conducive to achieving the complementarity of different features and improving the separability of different working mode categories in high-dimensional feature space.

[0078] The Pulse Repetition Interval (PRI) is defined as the arrival time difference between adjacent pulses and has good stability. Figure 4a is a schematic diagram showing a result before pulse cleaning according to an exemplary embodiment. Figure 4b FIG. 1 is a schematic diagram showing a result before pulse cleaning according to an exemplary embodiment.

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

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

[0081] Use the following formula to calculate the duty cycle sequence:

[0082]

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

[0084] RF and PRI are important parameters that are tightly coupled with the radar operating mode. For example, RF agile inter-pulse modulation is often used in radar systems to improve the radar's anti-interference capability, and group-variable PRI inter-pulse modulation is used to resolve distance ambiguity or speed ambiguity. It can be seen that the selection of RF and PRI inter-pulse modulation types is more closely related to the radar operating mode. In order to effectively distinguish the differences between different inter-pulse modulation types, different modulation types can be converted into certain feature vectors for characterization. Introducing the sequence {κ i}, its i-th element can be expressed as:

[0085]

[0086] Among them, △x i represents the first-order difference of the i-th element, x i RF i or PRI i , ε′ represents the corresponding threshold, ε′=0.01σ x , σ x represents the mean of the corresponding sequence. Further, the eigenvector V can be defined i as follows:

[0087]

[0088] Using the above process, we can obtain the eigenvector V of RF and PRI rf and V pri .

[0089] Figure 5 is a schematic diagram showing a PRI modulation feature according to an exemplary embodiment, such as Figure 5 As shown in the figure, the PRI modulation characteristics corresponding to different PRI modulation types are shown. It can be seen that the characteristics of different PRI pulse modulation types are significantly different, which is of great value for subsequent radar working modes.

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

[0091]

[0092] Wherein, x represents the parameter vector 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 to obtain multiple subsequences.

[0094] A part of the existing working mode categories is used as known categories, and another part is used as unknown categories. The number of known categories and the number of unknown categories can be preset. Multiple subsequences are divided into training data sets and test data sets.

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

[0096] Performing histogram processing on the original pulse sequence to obtain a histogram corresponding to the original pulse sequence;

[0097] Get the PRI value with the highest frequency in the histogram;

[0098] Get the upper and lower bounds based on the PRI value;

[0099] The original pulse sequence is screened according to the upper bound and the lower bound to obtain the preprocessed original pulse sequence.

[0100] It can be understood that the original pulse sequence is cleaned and preprocessed based on the PRI parameters. The specific steps are as follows: the PRI sequence of the original pulse sequence is processed by histogram, and the statistical interval is set to 1us; find the PRI value with the highest frequency in the histogram, and take 10% of the value as the lower bound and 200% as the upper bound; filter the original PRI pulse sequence according to the upper and lower bounds of the pulse PRI value, and when it is within the lower and upper bounds, retain the various parameters of the original pulse, and if it is greater than the upper bound or less than the lower bound, set the values ​​of the corresponding pulse parameters to 0. Pulse cleaning processing can suppress the influence of missed and wrong pulses on the radar working mode to a certain extent.

[0101] S102, calculating the distance between the feature vector to be identified and the feature center of a known category to obtain a first distance; wherein the feature center of the known category is obtained by inputting a training data set into a DA network for training.

[0102] S103: Obtain a recognition result of the sample to be recognized according to the first distance and the distance threshold.

[0103] Optionally, S104 includes:

[0104] Identify the to-be-identified sample corresponding to the first distance being greater than the distance threshold as an unknown working mode;

[0105] Identify the to-be-identified sample as a known working mode when the first distance is less than or equal to the corresponding one;

[0106] The identification result of the sample to be identified is obtained based on the identified unknown working mode and known working mode.

[0107] It can be understood that the first distance of 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 embodiment, the simulation conditions are specifically as follows: the experiment is conducted under different levels of missed pulses and false pulses, and the data of 13 working modes of the airborne multi-function radar, including TWS, TAS, VS, RWS, multiple target tracking (MTT), single target tracking (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 non-cooperative target identification (TID), are simulated respectively, and the parameter value ranges of different working modes are shown in Table 1. Among them, each working mode generates 120,000 pulses; the ratio of missed pulses and false pulses is set between 10% and 40%; the measurement error is achieved by adding Gaussian white noise with a mean of 0 and a variance of 2us to each pulse.

[0109] Table 1

[0110] Working 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 settings are as follows: The parameters of the repetition type recognition network used in the present 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 verification and testing respectively.

[0112] Table 2

[0113]

[0114]

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

[0116] Figure 7 is a schematic diagram of an experimental result according to an exemplary embodiment. Figure 7 As shown, Figure 7 This is the experimental result when there are only leaky pulses. It can be seen that as the number of leaky pulses gradually increases, the recognition results have declined to a certain extent. This is because the existence of leaky pulses destroys the parameter rules of different working mode sequences, resulting in the overlap of the extracted deep features in the high-dimensional space. However, the overall recognition results in all cases are better than 83%, which shows that the DA algorithm is not sensitive to the leaky pulse situation.

[0117] Figure 8 is another schematic diagram of experimental results according to an exemplary embodiment. Figure 8 As shown, Figure 8 The experimental results when there are only wrong pulses are shown. Overall, the recognition results in the case of wrong pulses are lower than those when there are only missing pulses. This is because when the pulse sequence contains wrong pulses, the timing rules of different working mode sequences are destroyed, and new error attribute parameter information is introduced, making the situation more complicated. However, the experimental results effectively prove that the DA algorithm has good generalization for the case of wrong pulses.

[0118] Fig. 9 is another schematic diagram of experimental results according to an exemplary embodiment. Fig. 9 As shown, Fig. 9 The experimental results for the cases of missing and wrong pulses are shown. It can be seen that when the ratio of missing and wrong pulses exceeds 20% at the same time, the overall result shows a significant decline. This is because the parameter rules of different working modes are seriously destroyed due to the simultaneous existence of wrong and missing pulses. However, due to the use of DA network, feature information of different dimensions can be fused to achieve a certain degree of complementarity, thereby ensuring the stability of the recognition effect. Therefore, when the ratio of missing and wrong pulses is as high as 40% at the same time, the overall recognition result remains at around 80%. The above results show that DA network has good generalization in complex environments.

[0119] Fig.10 is a block diagram of an open set recognition device for radar working mode of a dual-path autoencoder according to an exemplary embodiment. Fig.10 As shown, the radar operating mode open set identification device 1000 of the dual-path autoencoder 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 to be identified; wherein the DA network is trained according to the training data set and the mixed loss function; the training data set is obtained according to the original pulse sequence; the mixed loss function is constructed according to 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 feature center of the known category to obtain a first distance; wherein the feature center of the known category is obtained by inputting a training data set into the DA network for training;

[0122] The identification module 1003 is used to obtain the identification result of the sample to be identified according to the first distance and the distance threshold.

[0123] Regarding the device in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0124] The preferred embodiments of the present invention are described in detail above in conjunction with the accompanying drawings. However, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, a variety of simple modifications can be made to the technical solution of the present invention, and these simple modifications all belong to 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 further describe various possible combinations.

[0126] In addition, various embodiments of the present invention may be arbitrarily combined, and as long as they do not violate the concept of the present invention, they should also be regarded as the contents disclosed by the present invention.

Claims

1. A radar operating mode open set identification method for a dual-path autoencoder, characterized in that: The method comprises: Inputting the sample to be identified into the trained DA network to obtain the feature vector to be identified of the sample to be identified; wherein the DA network is trained according to the training data set and the mixed loss function; the training data set is obtained according to the original pulse sequence; the mixed loss function is constructed according to the cross entropy loss function, the center loss function and the reconstruction loss function; Calculating the distance between the feature vector to be identified and the feature center of a known category to obtain a first distance; wherein the feature center of the known category is obtained by inputting the training data set into the DA network for training; The recognition result of the sample to be recognized is obtained according to the first distance and the distance threshold.

2. The radar operating mode open set identification method of the dual-path autoencoder according to claim 1 is characterized in that: Before inputting the sample to be identified into the trained DA network to obtain the feature vector to be identified of the sample to be identified, the method further includes: Inputting the training data set 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 reached, thereby obtaining the trained DA network and the known category feature center.

3. The radar operating mode open set identification method of the dual-path autoencoder according to claim 2 is characterized in that: The hybrid loss function is expressed as follows: Loss H =Loss ce +αLoss cl +β(Loss re_s +Loss re_p ); 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 Represents the semantic sequence reconstruction loss, Loss re_p represents the original pulse sequence reconstruction loss, α represents the first hyperparameter, and β represents the second hyperparameter.

4. The radar operating mode open set identification method of the dual-path autoencoder according to claim 2 is characterized in that: Before inputting the training data set into the DA network to train the DA network, the method further includes: Extracting semantic information from the preprocessed original pulse sequence to obtain a semantic sequence; Normalizing the original pulse sequence and the semantic sequence respectively to obtain the normalized original pulse sequence and the semantic sequence; Performing sliding window processing on the normalized original pulse sequence and the semantic sequence to obtain multiple subsequences; The plurality of subsequences are divided into the training data set and the test data set.

5. The radar operating mode open set identification method of the dual-path autoencoder according to claim 4 is characterized in that: Before extracting semantic information from the preprocessed original pulse sequence to obtain a semantic sequence, the method further includes: Performing histogram processing on the original pulse sequence to obtain a histogram corresponding to the original pulse sequence; Obtaining the PRI value with the highest frequency in the histogram; Obtain upper and lower bounds based on the PRI value; The original pulse sequence is screened according to the upper bound and the lower bound to obtain the preprocessed original pulse sequence.

6. The radar operating mode open set identification method of the dual-path autoencoder according to claim 4 is characterized in that: The semantic sequence includes the duty cycle, RF inter-pulse modulation characteristics, and PRI modulation characteristics corresponding to the original pulse sequence.

7. The radar operating mode open set identification method of the dual-path autoencoder according to claim 1, characterized in that: The DA network includes an original pulse sequence feature extraction subnetwork and a semantic sequence feature extraction subnetwork.

8. The radar operating mode open set identification method of the dual-path autoencoder according to claim 1, characterized in that: The process of obtaining the distance threshold includes: Calculate and obtain a second distance from each training sample in the training data set to the center of the known category feature; The distance threshold is obtained according to the box plot method and the second distance corresponding to each training sample.

9. The radar operating mode open set identification method of the dual-path autoencoder according to claim 1, characterized in that: The obtaining the recognition result of the sample to be recognized according to the first distance and the distance threshold includes: Identify the sample to be identified corresponding to the first distance being greater than the distance threshold as an unknown working mode; Identify the first distance being less than or equal to the corresponding to-be-identified sample as a known working mode; The identification result of the sample to be identified is obtained according to the identified unknown working mode and the known working mode.

10. A dual-channel autoencoder radar operating mode open set identification device, characterized in that: The device comprises: A 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 to be identified; wherein the DA network is trained according to the training data set and the mixed loss function; the training data set is obtained according to the original pulse sequence; the mixed loss function is constructed according to the cross entropy loss function, the center loss function and the reconstruction loss function; A distance acquisition module is used to calculate the distance between the feature vector to be identified and the feature center of a known category to obtain a first distance; wherein the feature center of the known category is obtained by inputting the training data set into the DA network for training; The recognition module is used to obtain the recognition result of the sample to be recognized according to the first distance and the distance threshold.

Images