A zero-shot learning-based open-set recognition method for signals

Abstract

The present invention discloses a method for open-set recognition of signals based on zero-shot learning. A LSTM model is trained using a signal sample set with known classes to obtain a target model f for open-set recognition of signals. The signal sample set with known classes is input into the trained target model f, and the feature mean vector and similarity threshold of each class of the signal sample set with known classes are calculated to obtain a set of feature mean vectors and a set of similarity thresholds. The signal sample set to be recognized is recognized. Through zero-shot learning and by learning the features of signals of unknown classes, the present invention achieves the purpose of improving the correct classification probability and signal classification accuracy of open-set signals.

Description

Technical Field

[0001] This patent relates to the fields of artificial intelligence and signal recognition, and specifically relates to a signal open-set recognition method based on zero-shot learning. Background Art

[0002] Nowadays, deep learning has received great attention in both the academic and industrial fields. Since the performance of deep learning has been greatly improved compared to traditional algorithms, deep learning has been widely applied in various fields, such as machine translation, image recognition, unmanned driving, natural language processing, etc. However, in real-world recognition and classification tasks, due to various objective factors, it is usually difficult to collect training samples of all classes when training a recognizer or classifier. A more complex real-world situation is open-set recognition, where the training data provided during training is not complete, and samples of unknown classes can be submitted to the deep learning model during the testing process. This requires the classifier to not only accurately classify the trained classes but also effectively handle the untrained classes. The goal of open-set recognition is to classify known classes (trained classes) while identifying unknown classes (untrained classes) as unknown. The key challenge of open-set recognition is how to simultaneously reduce the empirical classification risk of trained classes being classified into unknown classes and the open-space risk of untrained unknown classes being classified into known classes. In the task of signal open-set recognition, one type of method uses traditional deep learning models. Using the Softmax confidence score as the criterion for open-set recognition is prone to overconfidence, and the Softmax confidence score output of the model for unknown class samples may also be very high, making it impossible to judge whether a sample belongs to an unknown class or a known class by setting a confidence threshold. To solve the above problems, Openmax was proposed. By extracting the output of the last fully connected layer of the model for EVT fitting and then performing Openmax calculation, the confidence score of the open-set category is increased, changing the original N-class classification of the model to N+1 classification. This method can effectively solve the recognition of irrelevant open-set images, but it has poor performance in the signal open-set recognition task.

[0003] Another type of method uses a deep learning model to extract sample features and does not directly use the Softmax confidence score as the criterion for open-set recognition. Instead, it uses the output of the fully connected layer as the mapping feature of the sample in the high-dimensional space. When training the model, the loss function is used to encourage the mapping spaces of different classes to move away from each other or to strengthen the model's feature extraction ability for samples. When evaluating a sample, first load the known sample dataset to obtain the average value of the mapping features of the class in the high-dimensional space, then extract the mapping features of the test sample in the high-dimensional space through the model, and compare the Euclidean distance between the test sample and the known class samples to judge the similarity between the test sample and the known samples. A threshold is set for the similarity of the samples, and samples with a similarity less than the threshold are judged as unknown samples.

[0004] From the perspective of classification methods, existing signal open-set recognition methods all adopt a threshold-based classification scheme. That is, the deep model uses empirically set thresholds to classify input samples into certain known classes, and the threshold plays a key role in the classifier's operation. However, the current selection of empirical thresholds depends on known class information while ignoring unknown class information, which may pose risks due to the lack of available information on unknown class samples, resulting in a large number of unknown classes being classified into known classes. In fact, since the sample feature data of unknown classes can usually be obtained from low-similarity samples classified by the classifier into unknown classes, we can make full use of them to reduce the risk of unknown class samples being classified into known classes by the deep learning model, and further improve the robustness of these methods for identifying unknown class samples. Summary of the Invention

[0005] The object of the present invention is to provide a signal open-set recognition method based on zero-shot learning in view of the deficiencies of the prior art.

[0006] The object of the present invention is achieved by the following technical solutions: A signal open-set recognition method based on zero-shot learning includes the following steps:

[0007] (1) Use a signal sample set with known classes to train an LSTM model to obtain a target model f for signal open-set recognition;

[0008] (2) Input the signal sample set with known classes into the trained target model f, calculate the feature mean vector and similarity threshold for each class of the signal sample set with known classes, and obtain a set of feature mean vectors and a set of similarity thresholds;

[0009] (3) Input any signal sample x in the signal sample set to be recognized t into the trained target model f, and calculate the third sample feature of this signal sample x; t According to the set of feature mean vectors and the set of similarity thresholds obtained in step (2), obtain the maximum third similarity of this signal sample x; t Take out the similarity threshold that is consistent with the class of the maximum third similarity of signal sample x t from the set of similarity thresholds and compare them to determine whether this signal sample x t is a known class;

[0010] If it is determined that this signal sample x t is a known class, then the next signal sample x t+1 repeats the above steps;

[0011] If it is determined that this signal sample x t is an unknown class, add a new class to the known classes, and update the set of similarity thresholds and the set of feature mean vectors; the next signal sample xt+1 Perform step (4);

[0012] (4) The next signal sample x t+1 According to the updated similarity threshold set and the updated feature mean vector set obtained in step (3), repeat step (3) to obtain the maximum third similarity of this signal sample x t+1 Take out the similarity threshold that is consistent with the maximum third similarity category of the signal sample x from the updated similarity threshold set and compare them to determine whether this signal sample x t+1 is of a known category; t+1

[0013] If it is determined that this signal sample x t+1 belongs to a known category and does not belong to the newly added category, then the next signal sample x t+2 Repeats the above steps;

[0014] If it is determined that this signal sample x t+1 belongs to a known category and belongs to the newly added category, then update the feature mean vector set; the next signal sample x t+2 Perform step (5);

[0015] If it is determined that this signal sample x t+1 belongs to an unknown category, then add a new category to the known categories, and update the similarity threshold set and the feature mean vector set; the next signal sample x t+2 Perform step (5);

[0016] (5) The next signal sample x t+2 According to the updated similarity threshold set and the updated feature mean vector set obtained in step (4), repeat step (4) until corresponding categories of all signal samples in the signal sample set to be recognized are predicted.

[0017] Furthermore, step (1) specifically includes the following sub-steps:

[0018] (1.1) Obtain the signal sample set D with known categories train : D train ={x 1,1 , x 1,2 , …, x j,i , …, x n,N}, j = 1, 2…, j, …n, i = 1, 2…, i, …N, n represents that the signal sample set D with known categories train includes n types of signal samples, N represents that the signal sample set D with known categories train each type includes N signal samples; x j,i represents the signal sample set D with known categories train ​The signal sample of the $i$-th item belonging to the $j$-th category;

[0019] (1.2) Use the signal sample set $D$ with known categories train Train the LSTM model: Set the hyperparameters for training: The number of training epochs of the LSTM model is epoch, the number of batch samples is $M$, and the learning rate is $\eta$; The loss function of the LSTM model is

[0020]

[0021] where $e$ j,i represents the first sample feature of the signal sample $x$ j,i obtained through the LSTM model; $S$ j,i,k represents the first similarity between the sample feature $e$ j,i and the signal center $c$ k ; $S$ j,i,j represents the first similarity between the sample feature $e$ j,i and the signal center $c$ j ;

[0022] Input the signal sample $x$ j,i into the LSTM model, and the output of the LSTM model is defined as $\psi(x$ j,i ). Perform L2 normalization on the sample features extracted by the LSTM model to obtain the first sample feature $e$ j,i of the signal sample $x$ j,i :

[0023]

[0024] Calculate the average of the first sample features of the signal samples of each category, and use the average value as the signal center of the signal samples of this category; The calculation formula for the signal center of any category is as follows:

[0025]

[0026] where $c$ k represents the signal center of the signal samples of the $k$-th category, $k = 1, 2, \ldots, k, \ldots, n$;

[0027] Subsequently, calculate the first similarity between the first sample feature $e$ j,i of the signal sample $x$ ji and any signal center $c$ k to obtain the similarity $S$ j,i,k , and the calculation formula of the said $S$ j,i,k is as follows:

[0028]

[0029] where and $b$ are learnable parameters,

[0030] (1.3) After the training is completed, the target model f for signal open-set recognition is obtained.

[0031] Furthermore, the step (2) specifically includes the following sub-steps:

[0032] (2.1) Input the signal sample x in the signal sample set D with known categories train into the target model f. The output of the target model f is defined as f(x j,i ). Perform L2 normalization on the sample features extracted by the target model f to obtain the second sample feature e′ j,i of the signal sample x j,i : j,i

[0033]

[0034] Calculate the average of the second sample features of the signal samples of the k-th category, and use the average value as the feature mean vector C k of the signal samples of the k-th category. The calculation formula is as follows:

[0035]

[0036] Repeat the above steps to calculate the feature mean vector of each category in the signal sample set D train with known categories, and obtain the feature mean vector set C: C = {C1, C2, …, C k , …, C n}; where C k represents the feature mean vector of the k-th category;

[0037] (2.2) Set the similarity sets of n categories: S1, S2 … S k … S n , where S k represents the similarity set of the k-th category; The similarity set of each category is initialized to be empty;

[0038] For any signal sample x train in the signal sample set D with known categories, calculate the second similarity S′ j,i between the second sample feature e′ j,i of this signal and any feature mean vector C k . The calculation formula of the second similarity S′ j,i,k is as follows: j,i,k

[0039]

[0040] where, a and b2 are learnable parameters,

[0041] Repeat the above steps to obtain the signal sample x j,i The second similarity of each class: S′ j,i,1 , S′ j,i,2 …S′ j,i,k …S′ j,i,n , where S′ j,i,k represents the second similarity of the k-th class of the signal sample x j,i ; Arrange S′ j,i,1 , S′ j,i,2 …S′ j,i,k …S′ j,i,n in descending order to obtain the maximum value S′ j,i,K , and take S′ j,i,K as the maximum second similarity of the signal sample x j,i , K = 1, 2, …, K, …, n; If K = j, update the similarity set S j of the j-th class: Record the maximum second similarity S′ j,i of the signal sample x j,i,K into the similarity set S j of the j-th class. If K≠j, do not update the similarity set S j of the j-th class;

[0042] (2.3) Repeat step (2.2) for all signal samples in the signal sample set D train with known classes, and update the similarity set S1 of the first class, the similarity set S2 of the second class … the similarity set S k … of the k-th class and the similarity set S n … of the n-th class;

[0043] (2.4) Calculate that the number of the maximum second similarities of the similarity set S k of the k-th class is h k , and arrange the similarity set S k of the k-th class in descending order, and delete the last 5%h k of the maximum second similarities to obtain the minimum value limit_S k , and take limit_S k as the similarity threshold of the k-th class of the signal sample set D train with known classes;

[0044] (2.5) Repeat step (2.4) to calculate the similarity threshold of each class of the signal sample set D train with known classes, and obtain the similarity threshold set S: S = {limit_S1, limit_S2, …, limit_S k , …, limit_Sn}; where, limit_S k represents the similarity threshold for the k-th class.

[0045] Further, the step (3) specifically includes the following sub-steps:

[0046] (3.1) Input any signal sample x in the signal sample set to be recognized t into the target model f to obtain the third sample feature of the signal sample x t The signal sample x represents the t-th signal sample in the signal sample set to be recognized, and the signal sample set to be recognized includes n types of signal samples with known categories and m types of signal samples with unknown categories; t

[0047] (3.2) Subsequently, calculate the third similarity S between the third sample feature e of the signal sample x t and each feature mean vector in the feature mean vector set C: S t t,1 t,2 t,k t,n t t k t,k t,1 t,2 t,k t,n t,H t,H t H The calculation formula for the third similarity S between the third sample feature e of the signal sample x and the k-th feature mean vector C

[0048]

[0049]

[0050]

[0051] t,1 t,2 t,k t,n Sort S t,H t,H t H t,H H t in descending order to obtain the maximum value S t,H H t t > limit_S H t t,H H t t H t t is predicted as the H-th class signal; if S t,H t,H H t t t,H t,H H t t t,H t,H H t t H H t t H H t t H H t t H H t t H H t t H H t t H H t t H H t t H H t t H H t t H H t t H H t t H H t t H H t t H H t t H H t t H H t t H H t t H H t t H H t t H H t t H H t t H H t t H H t t H H t t H H t t H H t t H H t t H <0Predicted as a signal of class H, repeat the above steps to predict the next signal sample x t+1 for its class;

[0052] When the signal sample x t is predicted as an out-of-set signal sample, add a new class: n + 1 to the known n classes, and predict the signal sample x t as class n + 1; update the similarity threshold set S: add the similarity threshold limit_S for class n + 1 n+1 and set limit_S n+1 = 0.75; and update the feature mean vector set C: add the feature mean vector C for class n + 1 n+1 and set C n+1 = e t ; then perform step (4) to predict the class of the next signal sample x t+1 for its class.

[0053] Furthermore, the specific step (4) is as follows: for the next signal sample x t+1 According to the updated similarity threshold set S and the updated feature mean vector set C obtained in step (3.2), repeat step (3.2) to obtain the maximum third similarity S t+1 of the signal sample x t+1,H′ , H' = 1, 2,..., H',..., n + 1; and take the similarity threshold limit_S of class H' from the updated similarity threshold set S H′ ;

[0054] If S t,H′ > limit_S H′ , then the signal sample x t+1 is predicted as a signal of class H'; if S t,H′ < limit_S H′ , then the signal sample x t+1 is predicted as an out-of-set signal sample;

[0055] When the signal sample x t+1 is predicted as a signal of class H' and H' ≤ n, repeat the above steps to predict the class of the next signal sample x t+2 for its class;

[0056] When the signal sample x t+1 is predicted as a signal of class H' and H' > n, update the feature mean vector set C: obtain the third sample feature e t+1 of the signal sample x t , and take the average of the third sample feature e t and the feature mean vector C H′ of class H' as the new feature mean vector C H′; Subsequently, step (5) is performed to predict the next signal sample x t+2 's category;

[0057] When the signal sample x t+1 is predicted as an out-of-set signal sample, a new category: n + 2 is added to the known n + 1 categories, and the signal sample x t+1 is predicted as the (n + 2)-th category; update the similarity threshold set S: add the similarity threshold limit_S for the (n + 2)-th category n+2 and set limit_S n+2 = 0.75; and update the feature mean vector set C: add the feature mean vector C for the (n + 2)-th category n+2 and set C n+1 = e t+1 ; Subsequently, step (5) is performed to predict the next signal sample x t+2 's category.

[0058] The beneficial effects of the present invention are as follows: In the application scenario of out-of-set recognition of signals by machine learning, the performance of out-of-set recognition is poor, and it is easy to classify unknown signals into known signals. Most current out-of-set recognition methods only use the information of known class samples to train the classifier, ignoring the additional information that unknown class samples can provide. The present invention achieves the purpose of improving the correct classification probability and classification accuracy of out-of-set signals through zero-shot learning by learning the features of unknown class signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 is a global schematic diagram of a signal out-of-set recognition method based on zero-shot learning;

[0060] Figure 2 is a flowchart for updating the center of known classes in the method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0061] In order to make the purpose, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts are within the protection scope of the present invention.

[0062] In view of the problems existing in signal data in open-set recognition, the purpose of the present invention is to provide a signal open-set recognition technology based on zero-shot learning. By classifying the unknown classes into known unknown classes and unknown unknown classes, the samples classified into the position classes by the deep learning network are extracted with high-dimensional space features as the sample centers of the known unknown classes, and the original N-classifier is changed into an N+M-class classifier, where M is the number of sample centers of the unknown classes, which can effectively reduce the probability of the unknown classes being classified into the known classes and improve the open-set recognition rate.

[0063] As Figure 1 and Figure 2 shown, the present invention provides a signal open-set recognition method based on zero-shot learning, including the following steps:

[0064] (1) In this embodiment, a signal sample set containing 11 classes with 1000 signal samples in each class is obtained from the RadioSignalRML2016_10a dataset; 80% of the signal samples in each class are extracted from the signal sample set as the total training set D train_All , and the remaining 20% of the signal samples are used as the test set D test : D test = {x1, x2, …, x t , …, x 2200}; 9 classes of signal samples are randomly selected from the total training set D train_All as the training set D train : D train = {x 1,1 , x 1,2 , …, x j,i , …, x 9,800}, where j = 1, 2 …, j, … 9, i = 1, 2 …, i, … 800, and x j,i represents the i-th signal sample in the j-th class in the training set D train . The training set D train includes n = 9 classes of signal samples and N = 800 signal samples in each class.

[0065] The RadioSignalRML2016_10a dataset is divided into 11 classes, specifically including 8 digital modulation methods: 8PSK, BPSK, CPFSK, GFSK, PAM4, 16QAM, 64QAM, QPSK and 3 analog modulation methods: AM-DSB, AM-SSB, WBFM.

[0066] (2) Use the signal sample set with known categories to train the LSTM model to obtain the target model f for signal open-set recognition;

[0067] The step (2) specifically includes the following sub-steps:

[0068] (2.1) Train the LSTM model using a set of signal samples with known categories. In this embodiment, the set of signal samples with known categories is the training set D train : D train ={x 1,1 , x 1,2 , …, x j,i , …, x 9,800}.

[0069] (2.2) Set the hyperparameters for training: the number of training epochs for the LSTM model is epoch, the number of batch samples is M, and the learning rate is η; the loss function of the LSTM model is

[0070]

[0071] where e j,i represents the first sample feature of the signal sample x j,i ; S j,i,k represents the first similarity between the sample feature e j,i and the signal center c k ; S j,i,j represents the first similarity between the sample feature e j,i and the signal center c j .

[0072] Input the signal sample x j,i into the LSTM model. The output of the LSTM model is defined as ψ(x j,i ). Perform L2 normalization on the sample features extracted by the LSTM model to obtain the first feature vector e j,i of the signal sample x j,i :

[0073]

[0074] Calculate the average of the first feature vectors of the signal samples for each category, and use the average value as the signal center of the signal samples of that category; the calculation formula for the signal center of any category is as follows:

[0075]

[0076] where c k represents the signal center of the signal samples of the k-th category, k = 1, 2 …, k, … 9.

[0077] Subsequently, calculate the similarity between the first sample feature e ji and any signal center c k to obtain the first similarity S j,i,k . The first similarity S j,i,kThe calculation formula is as follows:

[0078]

[0079] Wherein, and b1 are learnable parameters,

[0080] (2.3) After the training is completed, the target model f for signal open set recognition is obtained. The loss function L(e j,i ) is conducive to pushing the sample feature vector closer to the center of this class and pulling it away from the centers of all other classes.

[0081] (3) Input the signal sample set with known classes into the trained target model f, calculate the feature mean vector and similarity threshold for each class of the signal sample set with known classes, and obtain the feature mean vector set and similarity threshold set;

[0082] The step (3) specifically includes the following sub-steps:

[0083] (3.1) Input the signal sample x train in the training set D j,i into the target model f. The output of the target model f is defined as f(x j,i ). Perform L2 normalization on the sample features extracted by the target model f to obtain the second feature vector e′ j,i of the signal sample x j,i :

[0084]

[0085] Average the second feature vectors of the signal samples of the k-th class, and use the average value as the feature mean vector C k of the signal samples of the k-th class. The calculation formula is as follows:

[0086]

[0087] Repeat the above steps to calculate the feature mean vector of each class of the signal sample set D train with known classes, and obtain the feature mean vector set C: C = {C1, C2,..., C k ,..., C9}; where C k represents the feature mean vector of the k-th class;

[0088] (3.2) Set the similarity sets for 9 classes: S1, S2... S k ... S9, where S k represents the similarity set of the k-th class; The similarity set of each class is initialized to be empty;

[0089] For the training set D trainAny one of the signal samples x j,i , calculate the second eigenvector e′ of the signal j,i and any one of the eigen-mean vectors C k for the second similarity S′ j,i,k . The calculation formula for the second similarity S′ j,i,k is as follows:

[0090]

[0091] where and b2 are learnable parameters

[0092] Repeat the above steps to obtain the second similarity of each class of the signal sample x j,i : S′ j,i,1 , S′ j,i,2 … S′ j,i,k … S′ j,i,9 , where S′ j,i,k represents the second similarity of the k-th class of the signal sample x j,i . Arrange S′ j,i,1 , S′ j,i,2 … S′ j,i,k … S′ j,i,9 in descending order to obtain the maximum second similarity S′ j,i of the signal sample x j,i,K , K = 1, 2, …, K, …, 9; if K = j, update the similarity set S j of the j-th class: record the maximum second similarity S′ j,i of the signal sample x j,i,K into the similarity set S j of the j-th class. If K ≠ j, do not update the similarity set S j of the j-th class;

[0093] (3.3) Repeat step (3.2) for all the signal samples in the training set D train to update the similarity set S1 of the first class, the similarity set S2 of the second class … the similarity set S k … of the k-th class and the similarity set S9 of the ninth class;

[0094] (3.4) Calculate the number of the maximum second similarities of the similarity set S k of the k-th class as h k , and arrange the similarity set S k of the k-th class in descending order, and delete the maximum second similarities at the end 5%h k to obtain the minimum value limit_S k . Take limit_S k as the signal sample set D with known classestrain Similarity threshold for the k-th class;

[0095] (3.5) Repeat step (3.4) to calculate the signal sample set D with known classes train The similarity threshold for each class, to obtain the similarity threshold set S: S = {limit_S1, limit_S2, …, limit_S k , …, limit_S9}; where limit_S k represents the similarity threshold for the k-th class.

[0096] (4) Input any signal sample x in the signal sample set to be recognized t into the trained target model f, and calculate the third sample feature of this signal sample x t ; and according to the feature mean vector set and similarity threshold set obtained in step (2), obtain the maximum third similarity of this signal sample x t , and take out the similarity threshold that is consistent with the maximum third similarity class of the signal sample x t from the similarity threshold set and compare them to determine whether this signal sample x t is known in class;

[0097] If it is determined that this signal sample x t is known in class, then the next signal sample x t+1 repeats the above steps;

[0098] If it is determined that this signal sample x t is unknown in class, then add a new class to the known classes, and update the similarity threshold set and the feature mean vector set; the next signal sample x t+1 performs step (4).

[0099] The specific steps of step (4) include the following sub-steps:

[0100] (4.1) Input any signal sample x in the signal sample set to be recognized t into the target model f, to obtain the third sample feature e of the signal sample x t : t : The signal sample x t represents the t-th signal sample in the signal sample set to be recognized; in this implementation, the signal sample set to be recognized is the test set D test : D test = {x1, x2, …, x t , …, x 2200}; the test set D test includes 9 classes of signal samples with known classes and 2 classes of signal samples with unknown classes;

[0101] (4.2) Subsequently, calculate the third sample feature e of the signal sample x t and the third similarity of each feature mean vector in the set C of feature mean vectors: S t 、S t,1 、S t,2 …S t,k …S t,9 ; For the signal sample x t the third sample feature e t and the third similarity S k with the k-th feature mean vector C t,k is calculated as follows:

[0102]

[0103] Arrange S t,1 、S t,2 …S t,k …S t,9 in descending order to obtain the maximum value S t,H , and take S t,H as the maximum third similarity of the signal sample x t , where H = 1, 2, …, H, …, 9; Take the similarity threshold limit_S of the H-th class from the set of similarity thresholds H ;

[0104] If S t,H > limit_S H , then the signal sample x t is predicted as the signal of the H-th class; If S t,H < limit_S H , then the signal sample x t is predicted as an out-of-set signal sample;

[0105] When the signal sample x t is predicted as the signal of the H-th class, repeat the above steps to predict the class of the next signal sample x t+1 ;

[0106] When the signal sample x t is predicted as an out-of-set signal sample, add a new class: 10 to the known 9 classes, and predict the signal sample x t as the 10th class; Update the set S of similarity thresholds: Add the similarity threshold limit_S of the 10th class 10 and set limit_S 10 = 0.75; And update the set C of feature mean vectors: Add the feature mean vector C of the 10th class 10 and set C 10 = e t; Subsequently, step (5) is performed to predict the next signal sample x t+1 of the class.

[0107] (5) The next signal sample x t+1 According to the updated similarity threshold set and the updated feature mean vector set obtained in step (4), repeat step (3) to obtain the maximum third similarity of this signal sample x t+1 ; Take out the similarity threshold consistent with the maximum third similarity class of the signal sample x from the updated similarity threshold set t+1 and compare it to determine whether this signal sample x t+1 is of a known class;

[0108] If it is determined that this signal sample x t+1 belongs to a known class and does not belong to the newly added class, then the next signal sample x t+2 repeats the above steps;

[0109] If it is determined that this signal sample x t+1 belongs to a known class and belongs to the newly added class, then update the feature mean vector set; the next signal sample x t+2 performs step (6);

[0110] If it is determined that this signal sample x t+1 is of an unknown class, then add a new class to the known classes, and update the similarity threshold set and the feature mean vector set; the next signal sample x t+2 performs step (6).

[0111] The specific steps of step (5) are as follows: The next signal sample x t+1 According to the updated similarity threshold set S and the updated feature mean vector set C, repeat step (4.2) to obtain the maximum third similarity S t+1 of the signal sample x t+1,H′ , H′ = 1, 2, …, H′, …, 10; and take the similarity threshold limit_S of the H’th class from the updated similarity threshold set S H′ ;

[0112] If S t,H′ > limit_S H′ , then the signal sample x t+1 is predicted to be a signal of the H’th class;

[0113] If S t,H′ < limit_S H′ , then the signal sample x t+1 is predicted to be an out-of-set signal sample;

[0114] When the signal sample x t+1Is predicted as a signal of the H'-th class and H' ≤ 9, repeat the above steps to predict the next signal sample x t+2 for its class;

[0115] When the signal sample x t+1 is predicted as a signal of the H'-th class and H' > 9, update the set of feature mean vectors C: Obtain the third sample feature e t+1 of the signal sample x t , and use the average of the third sample feature e t and the feature mean vector C H′ of the H'-th class as the new feature mean vector C H′ ;

[0116] When the signal sample x t+1 is predicted as an out-of-set signal sample, add a new class: 11 to the known 10 classes, and predict the signal sample x t+1 as the 11th class; update the set of similarity threshold S: Add the similarity threshold limit_S 11 for the 11th class and set limit_S 11 = 0.75; and update the set of feature mean vectors C: Add the feature mean vector C 11 for the 11th class and set C 11 = e t+1 ; then perform step (6) to predict the class of the next signal sample x t+2 ;

[0117] (6) For the next signal sample x t+2 According to the updated set of similarity thresholds and the updated set of feature mean vectors in step (5), repeat step (5) until the classes of all signal samples in the signal sample set to be recognized are predicted accordingly.

[0118] In this embodiment, "zero-shot learning" is reflected in the stage of training the LSTM model using the signal sample set with known classes, only using the signal samples with known classes and not using the signal samples with unknown classes; in the stage of obtaining the "set of feature mean vectors and set of similarity thresholds", setting thresholds for the known classes; in the stage of "identifying the class of any signal sample in the signal sample set to be recognized", the signal samples with unknown classes can be discriminated and classified.

[0119] Evaluation metrics: The commonly used open-set accuracy, closed-set accuracy, and precision are used as evaluation metrics for open-set performance.

[0120] 1) Open-set accuracy: The probability that a sample with an unknown class is classified into the unknown class;

[0121]

[0122] Among them, TU represents the number of samples with unknown categories that are correctly classified into the unknown class, and FK represents the number of samples with known categories that are misclassified into the unknown class; the higher the open-set accuracy rate, the better the recognition performance of the open-set recognition algorithm for the unknown class.

[0123] 2) Closed-set accuracy rate: The probability that a sample with a known category is classified into a known class;

[0124]

[0125] Among them, TK represents the number of samples with known categories that are correctly classified into known classes, and FU represents the number of samples with unknown categories that are misclassified into known classes; the higher the closed-set accuracy rate, the better the recognition performance of the open-set recognition algorithm for known classes.

[0126] 3) Precision rate: The proportion of correctly judged samples among the samples judged as member samples;

[0127]

[0128] Among them, TP represents the number of samples correctly judged, TP = TU + TK; FP represents the number of samples misjudged, FP = FK + FU; the higher the precision rate, the better the recognition performance of the open-set recognition algorithm for all classes.

[0129] Set the 8PSK and GFSK class signals as open-set classes, and the remaining 9 class signals as known classes. Each class of signals contains 200 signal samples, which are input into the open-set recognition model proposed by the present invention, and the TUR, TKR, and Acc results are calculated. The specific results are shown in Table 1. It can be seen from Table 1 that at 18 dB, the open-set recognition rate proposed by this patent reaches 88%, and the closed-set recognition rate reaches 99.77%, and the recognition effect is good. Its performance decreases with the decrease of the signal-to-noise ratio because the signal features are covered by noise at low signal-to-noise ratios.

[0130] Table 1 Experimental results of open-set recognition of signals

[0131]

[0132] The main content of the present invention is aimed at the application scenario of open-set recognition of signals in machine learning, where the open-set recognition performance is poor and it is easy to classify unknown signals into known signals. Currently, most open-set recognition methods only use the information of known class samples to train the classifier, ignoring the additional information that unknown class samples can provide. The present invention achieves the purpose of improving the correct classification probability and signal classification precision rate of open-set signals through zero-shot learning and learning the features of unknown class signals.

[0133] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A signal open-set recognition method based on zero-shot learning, characterized in that, Including the following steps: (1) Training an LSTM model using a signal sample set with known categories to obtain a target model f for signal open-set recognition; (2) Inputting the signal sample set with known categories into the trained target model f, calculating the feature mean vector and similarity threshold for each category of the signal sample set with known categories, and obtaining a feature mean vector set and a similarity threshold set; (3) Input any signal sample \(x\) in the signal sample set to be recognized t into the trained target model \(f\), and calculate the third sample feature of the signal sample \(x\) t ; Based on the set of characteristic mean vectors and the set of similarity thresholds obtained in step (2), the maximum third similarity of the signal sample x t is obtained. A similarity threshold that is consistent with the category of the maximum third similarity of the signal sample x t is retrieved from the set of similarity thresholds and compared to determine whether the signal sample x t is of a known category; If it is determined that the signal sample x t has a known class, then for the next signal sample x t+1 repeat the above steps; If it is determined that the signal sample x t has an unknown category, a new category is added to the known categories, and the similarity threshold set and the feature mean vector set are updated; for the next signal sample x t+1 perform step (4); (4) The next signal sample x t+1 Based on the updated set of similarity thresholds and the updated set of feature mean vectors obtained in step (3), repeat step (3) to obtain the signal sample x t+1 of the maximum third similarity. Take out the similarity threshold that is consistent with the maximum third similarity category of the signal sample x from the updated set of similarity thresholds and compare them to determine whether the signal sample x t+1 is of a known category; t+1 ​ If it is determined that the signal sample x t+1 belongs to a known category and does not belong to a newly added category, then the next signal sample x t+2 repeats the above steps; If it is determined that the signal sample x t+1 belongs to a known category and is a new category, then update the set of feature mean vectors; for the next signal sample x t+2 perform step (5); If it is determined that the signal sample x t+1 belongs to an unknown category, a new category is added to the known categories, and the similarity threshold set and the feature mean vector set are updated; the next signal sample x t+2 performs step (5); (5) The next signal sample x t+2 According to the updated similarity threshold set and the updated feature mean vector set in step (4), repeat step (4) until corresponding categories are predicted for all signal samples in the signal sample set to be recognized.

2. The method for open-set signal recognition based on zero-shot learning according to claim 1, wherein The specific steps of step (1) include the following sub-steps: (1.1) Obtain a set of signal samples D with known categories train : D train = {x 1,1 , x 1,2 , …, x j,i , …, x n,N}, j = 1, 2 …, j, … n, i = 1, 2 …, i, … N, where n represents the number of categories of the set of signal samples D with known categories train includes n types of signal samples, and N represents the set of signal samples D with known categories train Each category includes N signal samples; x j,i represents the i-th signal sample in the j-th category of the set of signal samples D with known categories train ; (1.2) Use a known signal sample set D of the category train Train the LSTM model: Set the hyperparameters for training: the number of training epochs of the LSTM model is epoch, the number of batch samples is M, and the learning rate is η; the loss function of the LSTM model is Among them, e j,i represents the first sample feature of the signal sample x j,i ; S j,i,k represents the first similarity between the sample feature e j,i and the signal center c k ; S j,i,j represents the first similarity between the sample feature e j,i and the signal center c j ; Input the signal sample x j,i into the LSTM model. The output of the LSTM model is defined as ψ(x j,i ). Perform L2 normalization on the sample features extracted by the LSTM model to obtain the first sample feature e j,i of the signal sample x j,i : Taking the average of the first sample features of the signal samples of each category and using the average value as the signal center of the signal samples of this category; the calculation formula for the signal center of any category is as follows: Among them, c k represents the signal center of the signal samples of the k-th class, where k = 1, 2, …, k, …, n; Subsequently, calculate the first sample feature e j,i of the signal sample x ji and the first similarity with any signal center c k to obtain the similarity S j,i,k , and the calculation formula of the said S j,i,k is as follows: wherein, and b are learnable parameters, (1.3) After the training is completed, a target model f for signal open-set recognition is obtained.

3. The method for signal open-set recognition based on zero-shot learning according to claim 2, wherein The specific steps of step (2) include the following sub-steps: (2.1) Input the signal sample x in the signal sample set D with known categories train into the target model f. The output of the target model f is defined as f(x j,i ). Perform L2 normalization on the sample features extracted by the target model f to obtain the second sample feature e′ j,i of the signal sample x j,i : j,i ​ Average the second sample features of the signal samples of the k-th class, and use the average value as the feature mean vector C of the signal samples of the k-th class k , and the calculation formula is as follows: Repeat the above steps to calculate the set D of signal samples with known categories train of the feature mean vectors of each category, obtaining the set C of feature mean vectors: C = {C1, C2, …, C k , …, C n}; where C k represents the feature mean vector of the k-th category; (2.2) Set the similarity sets for n classes: S1, S2…S k …S n , where S k represents the similarity set of the k-th class; the similarity set of each class is initialized to be empty; For any signal sample x in the signal sample set D with known categories train calculate the second sample feature e′ of the signal j,i and the second similarity S′ with any feature mean vector C j,i The calculation formula of the second similarity S′ k is as follows: j,i,k The second similarity S′ j,i,k is as follows: wherein, and b2 are learnable parameters, Repeat the above steps to obtain the signal sample x j,i The second similarity of each class: S′ j,i,1 , S′ j,i,2 …S′ j,i,k …S′ j,i,n , where S′ j,i,k represents the second similarity of the k-th class of the signal sample x j,i ; Arrange S′ j,i,1 , S′ j,i,2 …S′ j,i,k …S′ j，i,n in descending order to obtain the maximum value S′ j,i,K , and take S′ j,i,K as the maximum second similarity of the signal sample x j,i , K = 1, 2, …, K, …, n; If K = j, update the similarity set S j of the j-th class: Record the maximum second similarity S′ j,i of the signal sample x j,i,K into the similarity set S j of the j-th class. If K ≠ j, do not update the similarity set S j ; (2.3) Repeat step (2.2) for all signal samples in the signal sample set D with known categories train to update the similarity set S1 of the first category, the similarity set S2 of the first category... the similarity set Sk of the k-th category k ... and the similarity set Sn of the n-th category n ; (2.4) Calculate the similarity set S of the k-th class k The maximum number of the second similarities is h k , and sort the similarity set S of the k-th class k in descending order, and delete the last 5%h k of the maximum second similarities to obtain the minimum value limit_S k , and use limit_S k as the signal sample set D with known classes train as the similarity threshold of the k-th class; (2.5) Repeat step (2.4) to calculate the signal sample set D with known categories train The similarity threshold for each category to obtain the similarity threshold set S: S = {limit_S1, limit_S2, …, limit_S k , …, limit_S n}; where limit_S k represents the similarity threshold for the k-th category.

4. The signal open-set recognition method based on zero-shot learning according to claim 3, wherein The specific steps of step (3) include the following sub-steps: (3.1) Input any signal sample \(x\) in the signal sample set to be recognized t into the target model \(f\) to obtain the third sample feature \(e\) of the signal sample \(x\) t t : The signal sample \(x\) t represents the \(t\)-th signal sample in the signal sample set to be recognized, and the signal sample set to be recognized includes \(n\) types of signal samples with known categories and \(m\) types of signal samples with unknown categories; (3.2) Subsequently, calculate the third sample feature e of the signal sample x t and the third similarity of each feature mean vector in the set C of feature mean vectors: S t , S t,1 ... S t,2 ... S t,k ... S t,n ; for the signal sample x t the third sample feature e t and the third similarity S k with the k-th feature mean vector C t,k is calculated according to the following formula: For S t,1 、S t,2 …S t,k …S t,n Arrange them in descending order, and the maximum value is S t,H . Take S t,H as the maximum third similarity of the signal sample x t , where H = 1, 2, …, H, …, n; take the similarity threshold limit_S of the H-th class from the similarity threshold set H ; If S t,H > limit_S H , then the signal sample x t is predicted as a signal of the H-th class; if S t,H < limit_S H , then the signal sample x t is predicted as an out-of-set signal sample; When the signal sample x t is predicted to be a signal of the H-th class, repeat the above steps to predict the class of the next signal sample x t+1 ; When the signal sample x t is predicted as an outlier signal sample, a new class: n + 1 is added to the known n classes, and the signal sample x t is predicted as the (n + 1)-th class; update the similarity threshold set S: add the similarity threshold limit_S for the (n + 1)-th class n+1 and set limit_S n+1 = 0.75; and update the feature mean vector set C: add the feature mean vector C for the (n + 1)-th class n+1 and set C n+1 = e t ; then perform step (4) to predict the class of the next signal sample x t+1 .

5. A method for open set recognition of signals based on zero-shot learning according to claim 4, characterized in that The specific content of step (4) is as follows: the next signal sample x t+1 According to the updated similarity threshold set S and the updated feature mean vector set C obtained in step (3.2), repeat step (3.2) to obtain the maximum third similarity S t+1 of the signal sample x t+1,H′ , where H′ = 1, 2, …, H′, …, n + 1; and take the similarity threshold limit_S of the H'-th class from the updated similarity threshold set S H′ ; If S t,H′ > limit_S H′ , then the signal sample x t+1 is predicted as a signal of the H'-th class; if S t,H′ < limit_S H′ , then the signal sample x t+1 is predicted as an out-of-set signal sample; When the signal sample x t+1 is predicted to be a signal of the H'-th class and H' ≤ n, repeat the above steps to predict the class of the next signal sample x t+2 ; When the signal sample x t+1 is predicted to be a signal of the H'-th class and H′ > n, update the set C of feature mean vectors: obtain the third sample feature e t+1 of the signal sample x t , and use the average of the third sample feature e t and the feature mean vector C H′ of the H'-th class as the new feature mean vector C H′ ; then proceed to step (5) to predict the class of the next signal sample x t+2 . When the signal sample x t+1 is predicted as an out-of-set signal sample, a new class: n + 2 is added to the known n + 1 classes, and the signal sample x t+1 is predicted as the (n + 2)-th class; update the similarity threshold set S: add the similarity threshold limit_S for the (n + 2)-th class n+2 and set limit_S n+2 = 0.75; and update the feature mean vector set C: add the feature mean vector C for the (n + 2)-th class n+2 and set C n+1 = e t+1 ; then perform step (5) to predict the class of the next signal sample x t+2 .

Images