A method for open set recognition of electromagnetic signals based on supervised contrastive learning

Through the method based on supervised comparison learning, combined with data enhancement and limit value theory, the problem of poor clustering of unknown types and embedded features in electromagnetic signal recognition is solved, and more efficient electromagnetic signal opening-set recognition performance is achieved.

CN115114958BActive Publication Date: 2024-12-0610TH RES INST OF CETC
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210739428.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-06-28
Publication Date
2024-12-06
Estimated Expiration
2042-06-28

AI Technical Summary

Technical Problem

When existing electromagnetic signal recognition methods face new radiation source types, radiation source individuals, modulation methods and encoding types, they cannot accurately identify unknown types, and the network model training efficiency is low and the clustering effect of known class embedding features is poor, resulting in Open set recognition performance is insufficient.

Method used

Using a method based on supervised contrast learning, the tail distribution of embedding features is fitted through data augmentation and limit value theory, deep neural network is trained, threshold thresholds between the embedding features and anchor points distances are set, joint inference recognition is realized, and model training efficiency is improved. Clustering effects of class sample embedding features are known.

Benefits of technology

It improves the accuracy and efficiency of electromagnetic signal open set recognition, enhances the ability to identify unknown classes, and at the same time improves the classification effect of known classes, reducing the risk of open space.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115114958B_ABST
    Figure CN115114958B_ABST
Patent Text Reader

Abstract

The present invention discloses an electromagnetic signal open set recognition method based on supervised contrastive learning, the method comprising the following steps: performing data enhancement on the electromagnetic signal; the electromagnetic signal comprises a continuous electromagnetic signal and a multi-pulse electromagnetic signal; using a supervised contrastive learning method to train a deep neural network, using the limit value theory to fit the tail distribution of the electromagnetic signal embedding feature; setting a threshold value of the distance between the embedded feature and its anchor point, and performing joint reasoning and recognition on the enhanced multiple electromagnetic signal copies, thereby obtaining a robust electromagnetic signal recognition result. The present invention fully considers the characteristics of different electromagnetic signals, extracts multiple copies of the steady-state part of the continuous signal, or multiple pulse signals as signal copies, trains a neural network based on a supervised contrastive loss function, improves the model training efficiency and the clustering effect of the embedded features of known class samples, and uses an integrated learning method in the reasoning stage to further improve the open set recognition effect of electromagnetic signals.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of electromagnetic signal recognition, in particular to an electromagnetic signal open set recognition method based on supervised contrast learning. Background Art

[0002] Electromagnetic signal identification includes radiation source type identification, radiation source individual identification, modulation type identification, channel coding identification, etc. It is an important prerequisite for subsequent radiation source identity authentication, signal demodulation and decoding. It plays a key role in military and civilian electromagnetic spectrum application fields such as signal monitoring, spectrum management, battlefield reconnaissance and electronic countermeasures.

[0003] Electromagnetic signal recognition methods can be mainly divided into likelihood-based methods, artificial feature-based methods and deep learning-based methods. Likelihood-based methods use hypothesis testing theory to obtain the best decision, but such methods usually require more prior knowledge and higher computational complexity. Artificial feature-based methods design and extract artificial features for classification. These features can be summarized as time domain features, statistical features, and change domain features. This method relies heavily on expert experience and performs well on specialized solutions, but has poor versatility. In recent years, considering the advantages of deep learning methods such as no need to design artificial features and high accuracy, some studies have applied deep learning methods to electromagnetic signal recognition and achieved remarkable results.

[0004] Although considerable achievements have been made in the research of electromagnetic signal recognition, the electromagnetic environment is becoming more and more complex, and new radiation source types, radiation source individuals, modulation methods, and coding types are constantly emerging. The known classical electromagnetic signal recognition algorithms and intelligent electromagnetic signal recognition algorithms based on deep learning can only recognize known types. When a new type of signal appears, the existing methods can only mistakenly identify it as one of the known types, and cannot accurately identify it as an unknown type. Therefore, the problem of open set recognition of electromagnetic signals needs to be solved urgently.

[0005] The open set recognition problem is defined as follows: the training phase contains K types of samples, and the inference phase contains N types of samples (N>K), including K known classes (the same as the training phase) and NK unknown classes. The goal is to identify the unknown classes and correctly classify the known classes. Because the unknown class signals are not involved in the training recognition model and there is no prior information about the unknown class signals, solving the open set recognition problem is very challenging.

[0006] At present, the research on open set recognition is mainly concentrated in the field of image recognition, and the main models can be divided into discriminative models and generative models (for example, see non-patent literature 1). Most discriminative models judge whether a sample belongs to a known class or an unknown class based on a set empirical threshold. However, the selection of empirical thresholds based on models trained with known class samples for identifying unknown classes will inevitably lead to open space risks. In order to reduce this risk, some studies are committed to making full use of known class data, such as using extreme value theory (EVT) to model the tail distribution of known data evaluation distribution and determine a better threshold. Generative models usually use generative models to construct false data to simulate unknown class samples, thereby converting the open set recognition problem into an N+1 class classification problem, where N represents the number of categories of known class samples and 1 represents other unknown types. Common generative models include generative adversarial networks (GAN) and autoencoders.

[0007] In the field of electromagnetic signal recognition, some scholars have studied the open set recognition problem of individual radiation sources and modulation modes. One method is to use special network models or loss functions to extract sample features in the training stage, enhance the clustering effect of known class sample features, and thus improve the recognition effect of unknown classes in the inference stage. The network models or loss functions involved include twin networks and triplet loss functions (for example, see non-patent literature 2 and non-patent literature 3). Another method is to refer to or improve the open set recognition method in the field of image recognition for electromagnetic signal open set recognition. Xu Yujie of the University of Science and Technology of China, based on the intra-class splitting algorithm (ICS), conducted in-depth research on the construction method of decision boundary samples and proposed an improved wireless signal open set recognition algorithm based on decision boundary samples. The effectiveness of the algorithm was verified on electromagnetic signal data sets and public image data sets (for example, see non-patent literature 4); Hao Yunfei et al. of the National University of Defense Technology proposed open set recognition of signal modulation modes based on generative adversarial networks. When the signal-to-noise ratio is greater than 0dB, the model can achieve better open set recognition effect of signal modulation modes (for example, see non-patent literature 5).

[0008] Through the investigation and analysis of open set recognition technology in the field of image recognition and electromagnetic signal recognition, the problems and defects of existing technologies are summarized:

[0009] (1) The characteristics of electromagnetic signals are not combined to improve the open set recognition performance. Electromagnetic signals can generally be divided into continuous signals and pulse signals. Multiple copies with the same steady-state characteristics are extracted from the steady-state part of the continuous signal. The same radiation source continuously sends multiple pulses, and each pulse is used as a different copy of a signal. Combined with the idea of ​​ensemble learning, the open set recognition performance of electromagnetic signals can be further improved.

[0010] (2) The given network model or loss function has low training efficiency, and the known class embedding feature clustering effect is poor. The contrast loss function and ternary loss function used by the twin network are trained in the form of paired samples or triplet samples, which has low training efficiency and poor embedding feature clustering effect.

[0011] Reference List

[0012] Non-patent literature

[0013] Non-patent literature 1: Geng C, Huang S, Chen S. Recent advances in open set recognition: A survey [J]. IEEE transactions on pattern analysis and machine intelligence, 2020, 43(10): 3614-3631.

[0014] Non-patent literature 2: Guo Youwei. Research on automatic modulation recognition technology based on deep learning[D]. China Academy of Engineering Physics, 2018.

[0015] Non-patent literature 3: Song Yuxuan. Research on evolutionary deep learning and application in communication signal recognition[D]. Xidian University, 2020.

[0016] Non-patent literature 4: Xu Yujie. Research on open set recognition technology of wireless signals based on deep learning[D]. University of Science and Technology of China, 2021.

[0017] Non-patent literature 5: Hao Yunfei, Liu Zhangmeng, Guo Fucheng, Zhang Min. Open set recognition of signal modulation based on generative adversarial network[J]. Systems Engineering and Electronics, 2019, 41(11): 2619-2624. Summary of the invention

[0018] In view of this, the present invention provides an electromagnetic signal open set recognition method based on supervised contrastive learning, which can improve the model training efficiency and the clustering effect of embedding features of known class samples, and further improve the electromagnetic signal open set recognition effect.

[0019] The present invention discloses an open set recognition method for electromagnetic signals based on supervised contrastive learning, comprising the following steps:

[0020] Step 1: performing data enhancement on the electromagnetic signal; the electromagnetic signal includes a continuous electromagnetic signal and a multi-pulse electromagnetic signal;

[0021] Step 2: Use supervised contrastive learning method to train the deep neural network and use limit value theory to fit the tail distribution of the embedded features of the electromagnetic signal;

[0022] Step 3: Set the threshold of the distance between the embedded feature and its anchor point, and perform joint reasoning and recognition on the enhanced multiple electromagnetic signal copies to obtain a robust electromagnetic signal recognition result.

[0023] Furthermore, the embedded features of electromagnetic signals are extracted via a supervised contrastive learning method, which has the characteristics of intra-class aggregation and inter-class separation.

[0024] Furthermore, for the multi-pulse electromagnetic signal, multiple electromagnetic pulse signals emitted by the same radiation source are obtained through signal sorting technology, and the multiple electromagnetic pulse signals are used as different viewing angle data of the same signal to achieve data enhancement.

[0025] Furthermore, for the continuous electromagnetic signal, signal slices containing the same steady-state characteristics are obtained through signal slicing technology, and multiple electromagnetic signal slices are used as different viewing angle data of the same signal to achieve data enhancement.

[0026] Furthermore, in step 2, N communication signals or N groups of multi-pulse signals are randomly selected from the training set, and K sub-signals are randomly intercepted from each communication signal, or K pulses are selected from each group of multi-pulse signals as K perspectives of the signal to form a batch of multi-perspective training sample sets. in represents the lth training sample vector, is the label of the lth training sample.

[0027] Furthermore, in step 2, the network architecture of the training phase consists of an encoding network and a mapping network; wherein the encoding network is represented as Enc(·), which is used to map the input signal x into a representation vector According to step 1, the electromagnetic signal samples are data enhanced, and the enhanced signal data from different perspectives are respectively input into the unified coding network to obtain their respective representation vectors; the representation vector r is normalized to The mapping network is represented by Proj(·), which is used to map the representation vector r to a vector Normalize the vector z to The unit hypersphere of space is on the surface of the unit hypersphere of space, which is convenient for mapping space Computes the vector inner product.

[0028] Furthermore, the encoding network can be any type of deep neural network; the deep neural network includes a residual network ResNet, a densely connected network DenseNet, and a network with parallel connections GoogleNet; the mapping network can be a multi-layer perceptron or a fully connected network containing only one linear layer.

[0029] Furthermore, the supervised contrast loss function is used to train the deep neural network, which is expressed as follows:

[0030]

[0031] Where i∈I≡{1,...,KN} represents the index of any multi-view signal. represents the supervised contrast loss function with multi-view signal i as the anchor point, a∈A(i)≡I\{i} represents the index of all signals except signal i in a batch of multi-view signals, represents the set of signals with the same label as signal i other than signal i in a batch of multi-view signals, z i represents the output vector of the multi-view signal i after passing through the encoding network and the mapping network, z p Represents the output vector of the signal with the same label as signal i after passing through the encoding network and mapping network, z a represents the output vector of other signals other than signal i after passing through the encoding network and the mapping network. The symbol · represents the inner product operation of the output vectors of the two mapping networks. The symbol | | represents the number of elements in the set. Represents a scalar temperature parameter, log is a logarithmic operation, and exp is an exponential operation.

[0032] Furthermore, in step 2, the samples in the training set are input into the trained neural network, the representation vector r of each sample is extracted, and the corresponding elements of the representation vectors of different categories are averaged to obtain the average representation vector of each category. As the anchor point, m represents the average, c∈{0,1,...,C} represents the category, and C represents the number of sample categories in the training set. According to the limit value theory, the distance between the representation vector of each category in the training sample and the average representation vector is fitted using Weibull distribution. The tail distribution of c Represents the output vector of a category c sample after the encoding network.

[0033] Furthermore, the cumulative probability density function of the Weibull distribution is expressed as:

[0034]

[0035] Here, η is the number of tail samples used to fit the Weibull distribution.

[0036] Furthermore, in step 3, in the inference stage, the tail distribution probability threshold p is set t , calculate the average distance from the representation vector of the multi-view sample to each type of anchor point Where k = 1, 2, ..., K represents the index of the inference signal copy, and K is the number of copies decomposed by an inference signal; the minimum average distance is used as a random variable to calculate the cumulative probability density of the Weibull distribution, and the cumulative probability density is compared with the tail distribution probability threshold p t If the comparison is less than the threshold, the signal corresponding to the group of copies is classified as the class with the minimum average distance from a certain class of anchor points. If it is greater than the threshold, it is identified as an unknown class, which is expressed as follows:

[0037]

[0038] in, Indicates taking the maximum d c The corresponding category index value, c=1,2,…,C, unknown represents the unknown class.

[0039] Due to the adoption of the above technical scheme, the present invention has the following advantages: the present invention fully considers the characteristics of different electromagnetic signals, extracts multiple copies of the steady-state part of the continuous signal, or multiple pulse signals as signal copies, trains a neural network based on a supervised contrast loss function, improves the model training efficiency and the clustering effect of the embedded features of known class samples, and adopts an integrated learning method in the inference stage to further improve the open set recognition effect of electromagnetic signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the embodiments of the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0041] Figure 1 A schematic diagram of a flow chart of an open set electromagnetic signal recognition method based on supervised contrastive learning according to an embodiment of the present invention;

[0042] Figure 2 Schematic diagram of a supervised contrastive learning training architecture according to an embodiment of the present invention. DETAILED DESCRIPTION

[0043] The present invention is further described in conjunction with the accompanying drawings and embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by ordinary technicians in this field should fall within the scope of protection of the embodiments of the present invention.

[0044] In view of the problems existing in the prior art, the present invention provides an open set identification method for electromagnetic signals, which is described in detail below in conjunction with the accompanying drawings.

[0045] Ordinary technicians in the industry can also use other steps to implement the electromagnetic signal open set identification method provided by the present invention. Figure 1 The electromagnetic signal open set identification method provided by the present invention is only a specific embodiment.

[0046] like Figure 1 As shown, the electromagnetic signal open set identification method provided by the embodiment of the present invention includes:

[0047] S1: performing data enhancement on an electromagnetic signal; the electromagnetic signal includes a continuous electromagnetic signal and a multi-pulse electromagnetic signal;

[0048] In S1 provided in the embodiment of the present invention, the electromagnetic signal sample data is enhanced, and the specific process is as follows:

[0049] Electromagnetic signals are usually divided into continuous signals and pulse signals. Continuous signals are mainly communication signals, such as Aircraft Communication Addressing and Reporting System (ACARS) signals, radio station signals, Wireless Local Area Networks (WLAN) signals, cellular network signals, etc. Continuous signals include the transient part when the radiation source is turned on and the steady-state part during the information transmission process. Different signal segments of the steady-state part have the same individual characteristics of the radiation source, modulation characteristics or channel coding characteristics; pulse signals include Automatic Dependent Surveillance-Broadcast (ADS-B) signals, radar pulse signals, etc. Pulse signal radiation sources usually emit multiple pulses in a short period of time. The receiver uses signal sorting technology to sort out the pulse signals emitted by the same radiation source to form a group of pulses. This group of pulses has the same individual characteristics of the radiation source or modulation characteristics.

[0050] Based on the above principle, the embodiment of the present invention randomly extracts a number of signal segments with the same number of sampling points from the steady-state part of the continuous signal to form a multi-view signal sample of the signal sample. Similarly, for the selected pulse group emitted by the same radiation source, multiple pulses are used as multi-view signal samples. The multi-view signal samples are used for subsequent neural network training and open set recognition reasoning, thereby accelerating the convergence of the neural network and improving the open set recognition performance.

[0051] S2: Use supervised contrastive learning to train deep neural networks and use limiting value theory to fit the tail distribution of embedded features of electromagnetic signals;

[0052] S3: Set the threshold of the distance between the embedded feature and its anchor point, and perform joint reasoning and recognition on the enhanced multiple electromagnetic signal copies to obtain a robust electromagnetic signal recognition result.

[0053] In this embodiment, the embedded features of the electromagnetic signal are extracted by a supervised contrastive learning method, which has the characteristics of intra-class aggregation and inter-class separation.

[0054] In this embodiment, for multi-pulse electromagnetic signals, multiple electromagnetic pulse signals emitted by the same radiation source are obtained through signal sorting technology, and the multiple electromagnetic pulse signals are used as different perspective data of the same signal to achieve data enhancement.

[0055] In this embodiment, for a continuous electromagnetic signal, a signal slice containing the same steady-state characteristics is obtained through a signal slicing technique, and multiple electromagnetic signal slices are used as different viewing angle data of the same signal to achieve data enhancement.

[0056] In this embodiment, in S2, N communication signals or N groups of multi-pulse signals are randomly selected from the training set, and K sub-signals are randomly intercepted from each communication signal, or K pulses are selected from each group of multi-pulse signals as K perspectives of the signal to form a batch of multi-perspective training sample sets. in represents the lth training sample vector, is the label of the lth training sample.

[0057] In this embodiment, in S2, see Figure 2 The network architecture in the training phase consists of an encoding network and a mapping network; the encoding network is represented as Enc(·), which is used to map the input signal x into a representation vector According to S1, the electromagnetic signal samples are data enhanced, and the signal data from different perspectives are input into the unified coding network to obtain their respective representation vectors; the representation vector r is normalized to The mapping network is represented by Proj(·), which is used to map the representation vector r to a vector Normalize the vector z to The unit hypersphere of space is on the surface of the unit hypersphere of space, which is convenient for mapping space Computes the vector inner product.

[0058] In this embodiment, the encoding network can be any type of deep neural network; deep neural networks include residual networks ResNet, densely connected networks DenseNet, and networks with parallel connections GoogleNet; the mapping network can be a multi-layer perceptron or a fully connected network containing only one linear layer.

[0059] In this embodiment, the supervised contrast loss function is used to train the deep neural network, which is expressed as follows:

[0060]

[0061] Where i∈I≡{1,...,KN} represents the index of any multi-view signal. represents the supervised contrast loss function with multi-view signal i as the anchor point, a∈A(i)≡I\{i} represents the index of all signals except signal i in a batch of multi-view signals, represents the set of signals with the same label as signal i other than signal i in a batch of multi-view signals, z i represents the output vector of the multi-view signal i after passing through the encoding network and the mapping network, z p Represents the output vector of the signal with the same label as signal i after passing through the encoding network and mapping network, z a represents the output vector of other signals other than signal i after passing through the encoding network and the mapping network. The symbol · represents the inner product operation of the output vectors of the two mapping networks. The symbol | | represents the number of elements in the set. Represents a scalar temperature parameter, log is a logarithmic operation, and exp is an exponential operation.

[0062] In this embodiment, in S2, the samples in the training set are input into the trained neural network, the representation vector r of each sample is extracted, and the corresponding elements of the representation vectors of different categories are averaged to obtain the average representation vector of each category. As the anchor point, m represents the average, c∈{0,1,...,C} represents the category, and C represents the number of sample categories in the training set. According to the limit value theory, the Weibull distribution is used to fit the distance between the representation vector of each category in the training sample and the average representation vector. The tail distribution of c Represents the output vector of a category c sample after the encoding network.

[0063] In this embodiment, the cumulative probability density function of the Weibull distribution is expressed as:

[0064]

[0065] Here, η is the number of tail samples used to fit the Weibull distribution.

[0066] In this embodiment, in S3, in the inference stage, the tail distribution probability threshold p is set t , calculate the average distance from the representation vector of the multi-view sample to each type of anchor point Where k = 1, 2, ..., K represents the index of the inference signal copy, and K is the number of copies decomposed by an inference signal; the minimum average distance is used as a random variable to calculate the cumulative probability density of the Weibull distribution, and the cumulative probability density is compared with the tail distribution probability threshold p t If the comparison is less than the threshold, the signal corresponding to the group of copies is classified as the class with the minimum average distance from a certain class of anchor points. If it is greater than the threshold, it is identified as an unknown class, which is expressed as follows:

[0067]

[0068] in, Indicates taking the maximum d c The corresponding category index value, c=1,2,…,C, unknown represents the unknown class.

[0069] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for open set recognition of electromagnetic signals based on supervised contrastive learning, characterized in that: The following steps are involved: Step 1: performing data enhancement on the electromagnetic signal; the electromagnetic signal includes a continuous electromagnetic signal and a multi-pulse electromagnetic signal; Step 2: Use supervised contrastive learning method to train the deep neural network and use limit value theory to fit the tail distribution of the embedded features of the electromagnetic signal; Step 3: Set the threshold of the distance between the embedded feature and its anchor point, and perform joint reasoning and recognition on the enhanced multiple electromagnetic signal copies to obtain a robust electromagnetic signal recognition result; In step 2, the network architecture of the training phase consists of an encoding network and a mapping network; wherein the encoding network is represented as , used to convert the input signal Mapped into a representation vector ; According to step 1, the electromagnetic signal sample is data enhanced, and the signal data of different perspectives after enhancement are respectively input into the unified coding network to obtain their respective representation vectors; the representation vector Normalized to On the unit hypersphere of space; the mapping network is expressed as , used to represent the vector Mapped to a vector ; The vector Normalized to The unit hypersphere of space is on the surface of the unit hypersphere of space, which is convenient for mapping space Calculate the vector inner product above; The supervised contrast loss function is used to train the deep neural network, which is expressed as follows: in, Represents the index of any multi-view signal, Indicates multi-view signal is the supervised contrast loss function of the anchor point, Indicates a batch of multi-view signals except for the signal The index of all signals of Represents a batch of multi-view signals In addition to the signal A collection of signals with the same label, Represents multi-view signal The output vector after encoding network and mapping network is: Representation and Signal The output vector of the signal with the same label after passing through the encoding network and the mapping network, Indicates signal The output vector of other signals after passing through the encoding network and mapping network is symbol Indicates the inner product operation of two mapping network output vectors, symbol Indicates the number of elements in a collection. Represents a scalar temperature parameter, log is a logarithmic operation, and exp is an exponential operation; The encoding network is any type of deep neural network; the mapping network is a multi-layer perceptron or a fully connected network containing only one linear layer.

2. The method according to claim 1, characterized in that The embedded features of electromagnetic signals are extracted by a supervised contrastive learning method, which has the characteristics of intra-class aggregation and inter-class separation.

3. The method according to claim 1, characterized in that For the multi-pulse electromagnetic signal, multiple electromagnetic pulse signals emitted by the same radiation source are obtained through signal sorting technology, and the multiple electromagnetic pulse signals are used as different viewing angle data of the same signal to achieve data enhancement.

4. The method according to claim 1, characterized in that For the continuous electromagnetic signal, signal slices containing the same steady-state characteristics are obtained through signal slicing technology, and multiple electromagnetic signal slices are used as different viewing angle data of the same signal to achieve data enhancement.

5. The method according to claim 1, characterized in that In step 2, randomly select Communication signal or Group multiple pulse signals and randomly intercept each communication signal sub-signals, or select from each group of multi-pulse signals Pulse as signal perspectives, forming a batch of multi-perspective training sample sets ,in Indicates training sample vectors, For the The labels of the training samples.

6. The method according to claim 1, characterized in that The deep neural network includes a residual network ResNet, a densely connected network DenseNet, and a network with parallel connections GoogleNet.

7. The method according to claim 1, characterized in that In step 2, the samples in the training set are input into the trained neural network to extract the representation vector of each sample. , average the corresponding elements of the representation vectors of different categories to obtain the average representation vector of each category As an anchor point, Indicates average, Indicates category, Represents the number of sample categories in the training set; according to the limit value theory, the Weibull distribution is used to fit the distance between the representation vector of each category in the training sample and the average representation vector The tail distribution of Represents a category The output vector of the sample after passing through the encoding network.

8. The method according to claim 7, characterized in that The cumulative probability density function of the Weibull distribution is expressed as: in, Represents the number of tail samples used to fit the Weibull distribution.

9. The method according to claim 1, characterized in that: In step 3, in the inference phase, the tail distribution probability threshold is set , calculate the average distance from the representation vector of the multi-view sample to each type of anchor point ,in represents the inference signal replica index, is the number of copies decomposed into an inference signal; the minimum average distance is used as a random variable to calculate the cumulative probability density of Weibull distribution, and the cumulative probability density is compared with the tail distribution probability threshold If the comparison is less than the threshold, the signal corresponding to the group of copies is classified as the class with the minimum average distance from a certain class of anchor points. If it is greater than the threshold, it is identified as an unknown class, which is expressed as follows: in, Indicates the maximum The corresponding category index value, , Indicates an unknown class.