An electromagnetic signal open set identification method based on data negative enhancement and pseudo-class prototype points

Abstract

The application discloses an electromagnetic signal open set identification method based on data negative enhancement and pseudo-class prototype points, and comprises the following steps: S1, dividing an original data set into a training set and a test set in proportion; S2, determining the values of class prototype points of K known classes used for training and a pseudo-class prototype point, and determining the values of other parameters when training a model; S3, inputting each sample into a neural network encoder to obtain a feature vector, and calculating the distance between the feature vector and each prototype point to obtain a distance vector; S4, obtaining the feature vector and the distance vector, and then selecting effective negative enhancement samples by using similarity screening; S5, inputting the feature vector and the distance vector obtained before into a loss function to calculate a loss, and updating the model; and S6, after the training is completed, calculating class prototype points used for testing and a threshold value used for discrimination. The application improves the open world adaptability of the model.

Description

TECHNICAL FIELD

[0001] The present application relates to prototype learning, data augmentation, electromagnetic modulation signal recognition, and in particular to a method for reducing the open world risk of a model by combining data negative augmentation with pseudo-class prototype points. BACKGROUND

[0002] With the great progress of deep learning in the field of pattern recognition, open world electromagnetic signal open set recognition has attracted more and more attention. Electromagnetic signal open set recognition refers to training an electromagnetic signal classification model through machine learning methods to enable it to reject unknown class samples from the open world, thereby improving the generalization ability of the classification model. The current open set recognition methods mainly target the image field, and there are few methods targeting the electromagnetic signal field. The open set recognition methods in the image field mainly include discriminative methods and generative methods. Discriminative methods use the characteristics of the classifier itself to identify unknown class samples, and generative methods use additional generated data to expand the function of the original closed set classifier. These methods work well in the image field, but cannot be perfectly applied to the electromagnetic signal field. Currently, electromagnetic signal open set recognition faces two major problems: one is that the classifier cannot effectively identify unknown inputs; the other is that the classifier misclassifies known class samples as unknown classes. We refer to these two major problems as open world risk.

[0003] To effectively address the open world risk, the present application uses data negative augmentation combined with class prototype learning to improve the open set recognition ability of the electromagnetic signal classification model. The core idea is to add pseudo-class prototype points to expand the open world adaptability of the model on the basis of traditional class prototype learning, and use the mechanism of data negative augmentation to enhance the adaptability of the model. SUMMARY

[0004] To overcome the problem that there are few existing electromagnetic signal open set recognition methods to address the open world risk, the present application provides an electromagnetic signal open set recognition method based on data negative augmentation and pseudo-class prototype points.

[0005] The present application is based on prototype learning, and additional pseudo-class prototype points are constructed to expand the feature space and preliminarily enhance the open world adaptability of the model. The method uses a random discarding negative augmentation method to enhance the data, selects appropriate augmented samples through similarity screening, and then updates the model parameters to enable the classifier to correctly guide the augmented samples to the region where the pseudo-class prototype points are located, further enhancing the open world adaptability of the model. In the case of ensuring that the closed set classification accuracy is not affected, the ability of the enhanced model to identify unknown inputs is enhanced. The training framework of the entire method is shown in the Figure 1

[0006] The technical solution adopted by the present application to solve its technical problems is: ​

[0007] An electromagnetic signal open-set recognition method based on data negative augmentation and pseudo-class prototype points, comprising the following steps:

[0008] S1: define the problem background and model of the method. The given training set composed of modulated signals is represented as where the input sample is a piece of electromagnetic modulated signal, L is the length of the signal, 2 represents the I and Q paths of the signal; y i ∈Y={1,2,...,K} is the class of the signal, K represents the number of known classes. Then, the test set of electromagnetic signals is represented as where Note that there is a K+1 class in the test set, which is called the unknown class, and all unknown inputs in the test stage are regarded as the K+1 class. We define the open-set recognition model as is the hypothesis space, X is the space where all input samples are located. Finally, the goal of electromagnetic signal open-set recognition is to reduce the following empirical risk:

[0009]

[0010] where, represents the expectation, represents the indicator function.

[0011] S2: initialize the class prototype points of the model of the method. In order to improve the robustness of the distribution of the feature space, we determine the position of the class prototype of each class at the initialization and keep it unchanged during the training stage. We call the class prototype points constructed in this way class anchor points. Using this idea, we preset a class prototype point for each known class. In order to increase the ability of the model to cope with the open world risk, an additional prototype point is added, which is called a pseudo-class prototype point. Thus, the class prototype point of the kth class (the prototype point of the K+1th class is regarded as a pseudo-prototype point) is represented as and,

[0012]

[0013] where α is a hyperparameter used to determine the position of the initial class prototype point. We do not update all class prototype points during the training process, but after the training is completed, we will update the class prototype points of the known classes again using the training samples, and keep the value of the pseudo-class prototype point unchanged.

[0014] S3: encode the input sample to obtain the feature vector and distance vector. The neural network encoder of the entire model is represented as where represents the feature space of the open-set recognition model, d z is the dimension of the feature space. By inputting the sample into The feature vector of the sample in the feature space can be obtained The Euclidean distance is used as the distance measurement method, and the distance calculation function is denoted as ξ. Then the feature vector of the sample can be calculated as The distance vector d composed of the distances between the sample and each class anchor point is calculated, where:

[0015]

[0016] Then the distance vector is normalized using the softmin function as follows:

[0017]

[0018] S4: At the same time, the original input sample is enhanced, and the feature vector and the distance vector are obtained after encoding. The purpose of using data enhancement is to further adjust the distribution of the feature space, especially the distribution of potential unknown inputs, using new data. We denote the data enhancer as ψ, and the enhanced data as x A = ψ(x). This method enhances the sample by random dropping, and the idea is to copy the original sample as follows:

[0019] x A = x (5)

[0020] Then, the number of dropped points is set as δ. δ non-repeated values are randomly selected in [0, L-1], and an index sequence idx is constructed. Then x is updated as follows: A

[0021]

[0022] The newly generated sample is used to update the model after similarity screening. The specific process is as follows: first, the negative enhanced sample is encoded by the encoder, and the distance vector is calculated:

[0023]

[0024] Then the distance vector is normalized using the softmin function, and the sample that meets the following rules is screened:

[0025]

[0026] Where y is the class of the sample corresponding to d A [:K] represents the slicing operation, that is, the first K elements are taken to form a new vector; β min and β max is a hyperparameter that controls the similarity. The purpose of using the slicing operation is to shield the pseudo prototype points, so as to calculate whether the feature vector of the enhanced sample maintains a certain similarity with the class prototype corresponding thereto.​

[0027] S5: Update the model parameters using the improved CAC distance loss function. For the samples from known classes, their preliminary loss function can be calculated as follows:

[0028]

[0029] For the augmented samples reserved by the similarity screening, their preliminary loss function can be calculated as follows:

[0030]

[0031] In addition, in order to minimize the distance between classes, a distance loss function constraint term is added as follows:

[0032]

[0033] Finally, the loss of each sample is calculated as follows:

[0034]

[0035] Where p and l are adjustable hyperparameters. Both of them are set to 0.1 in this method.

[0036] S6: After the training is completed, the class prototype points and the threshold for discrimination can be calculated for testing. First, after the training is completed, for each known class, the average value of the feature vectors of all training samples of the known class is taken as the class prototype of the class. Secondly, we select the average value of mind of all correctly identified known class samples as the basic threshold 0. Finally, during testing, for a test sample x, after encoding by the encoder and obtaining the distance vector d, the following decision method is used to judge the sample:

[0037]

[0038] The test framework of the whole method is shown in the accompanying drawing Figure 2 .

[0039] The working principle of this invention is as follows: First, an additional class prototype point is added to the traditional prototype learning model. Since this class prototype point does not participate in subsequent testing, we call it a pseudo-class prototype point. The pseudo-class prototype point reserves a new region in the feature space for the unknown class, so that during the testing phase, when encountering an unknown input, the trained classifier can map the unknown input to this new region, rather than the region where the known class is located. Second, simply setting pseudo-class prototype points is not enough, because if the pseudo-class prototype points are not involved in the model's training, the model will not be aware of their existence during the testing phase. Therefore, negative data augmentation is used to generate a batch of new samples, and the classifier is made to consider these data as unknown class samples, allowing these samples to participate in training along with the pseudo-class prototype points. The negatively generated samples need to satisfy the condition of having a certain similarity to the original samples but being identified as an unknown class. During testing, if an unknown input has some features of a known class, the classifier can also identify it as an unknown class as much as possible.

[0040] The advantages of this invention are: it improves the open-world adaptability of the model. By utilizing negative data augmentation and pseudo-class prototype points, it expands the open-set adaptability of the closed-set classifier. Negative data augmentation strengthens the classifier's discriminative ability, while the additional pseudo-class prototype points create an extra region in the feature space. Since negatively augmented samples and their corresponding known class samples still retain a certain similarity, and the classifier identifies them as unknown classes, it brings their feature vectors closer to the pseudo-class prototype points. This means that during the testing phase, when an input unknown class sample possesses certain characteristics of a known class, the classifier will tend to identify it as an unknown class. Attached Figure Description

[0041] Figure 1 This is a training block diagram for the electromagnetic signal open set recognition model based on negative data enhancement and pseudo-class prototype points proposed in this invention.

[0042] Figure 2 This is the test framework for the electromagnetic signal open set recognition model based on negative data enhancement and pseudo-class prototype points proposed in this invention. Detailed Implementation

[0043] The following is in conjunction with the instruction manual appendix. Figure 1 and attached Figure 2 The invention will be further described in detail using the modulation signal dataset RML2016.10A as an example.

[0044] The electromagnetic signal open set identification method based on negative data enhancement and pseudo-class prototype points according to the present invention specifically includes the following steps:

[0045] Step 1: Divide the dataset into two parts: a training set and a test set.

[0046] Step 2: Initialize the model and class prototype points.

[0047] Step 3: Encode the original input sample using the encoder and calculate the distance vector.

[0048] Step 4: Input the original input sample into the enhancer to obtain the negative enhanced sample, and then input the negative enhanced sample into the encoder to calculate the distance vector.

[0049] Step 5: Update the model parameters using the loss function.

[0050] Step 6: Update part of the model parameters, and finally evaluate the model performance with the test set.

[0051] In step 1, the specific operation process is as follows: for a complete RML2016.10A data set, there are 11 modulation types. First, randomly select 9 classes as known classes, and then the remaining 2 classes as unknown classes, according to which the entire data set is divided into two parts. Then, for each part, the training set and test set are divided according to the ratio of 7:3. Each input sample of the data set is which is a segment of modulation signal, where 2 represents I and Q, and 512 represents the number of sampling points.

[0052] In step 2, the specific operation process is as follows: the encoder of the entire model is composed of a 1-dimensional residual network, a linear layer, and 10 class prototype points. The residual network is composed of 3 residual blocks. The 10 class prototype points include 9 known class prototype points and 1 pseudo class prototype point. These class prototype points do not participate in the update of the encoder. From the attached Figure 1 figure, we can see that our model has two data processing paths, the upper path (solid line) processes the original sample, and the lower path (dashed line) processes the negative enhanced sample. When initializing the class prototype points, we set the parameter α that controls the position of the class prototype points to 5.

[0053] In step 3, the specific operation process is as follows: when training the model, we use the Adam optimizer and set the learning rate to 0.001. We use mini-batch training to train the model, and each batch contains 120 samples. At the beginning of each training round, a batch of samples is input into the model through the upper path, and then the feature vector is obtained and the distance vector is calculated.

[0054] In step 4, the specific operation process is as follows: at the beginning of each training round, the same batch of samples used in step 3 is input into the data enhancer through the lower path. Specifically, this method uses random dropout to obtain enhanced samples. For RML2016.10A samples, we randomly discard 75 sampling points. Then the enhanced sample is input into the encoder to obtain the output feature vector, and then the effective feature vector and distance vector are selected through similarity screening. We set the parameter β min and β max0.25 and 0.45, respectively.

[0055] In step 5, the specific operation process is as follows: as shown in the accompanying drawings Figure 1 As shown, the feature vectors and distance vectors obtained in steps 3 and 4 are used to calculate the loss. The feature vectors and distance vectors obtained in steps 3 and 4 are respectively input into the corresponding loss function according to formula (12). Finally, the gradients are calculated after accumulating all the losses, and the model parameters are updated.

[0056] In step 6, the specific operation process is as follows: for each known class, input the training set samples under the known class into the trained encoder to obtain the output feature vectors and the prediction results, accumulate the feature vectors with correct prediction results and take the average to obtain the class prototype point of the known class. Then calculate the average of the distance from each correctly identified feature vector to its class prototype point as the average minimum distance of the known class. Sum all the average minimum distances of the known classes and take the average to obtain the basic distance threshold θ. In the test stage, for a test sample, calculate the distance vector after calculating the distance from the feature vector obtained after inputting the test sample into the encoder to the class prototype point of each known class, and then use the given decision method to make a decision.

[0057] The present application adjusts the distribution of the feature space of the electromagnetic signal classification model based on data negative enhancement and pseudo prototype points, improves the discrimination ability of the model, and deals with the unknown input that may appear in the test. When unknown input appears, the classification model can identify it as unknown class as much as possible. The present application is merely illustrative, not restrictive. The skilled person understands that many changes, modifications, and even equivalents can be made within the spirit and scope of the invention defined in the claims, but all will fall within the protection scope of the present application.

Claims

1. An electromagnetic signal open set identification method based on data negative enhancement and pseudo-class prototype points, comprising the following steps: S1: define the problem background and the model; S2: initialize the class prototype points of the model; S3: encode the input sample to obtain the feature vector and the distance vector; S4: enhance the original input sample and obtain the feature vector and the distance vector after encoding; S5: update the model parameters using the improved CAC distance loss function; S6: after the training is completed, calculate the class prototype points and the threshold for testing; Step S1 specifically includes: representing a given training set consisting of modulated signals as where input samples are a piece of electromagnetic modulated signal, L is the length of the signal, and 2 indicates the I and Q paths of the signal; is the class of the signal, K indicates the number of known classes; then, an electromagnetic signal test set is represented as where ; note that there is an additional class in the test set , which is called the unknown class, and all unknown inputs in the test phase are considered as the class ; the open set recognition model is defined as , , is the hypothesis space, is the space where all input samples are located; finally, the goal of the electromagnetic signal open set recognition is to reduce the following empirical risk: (1) wherein denotes the expectation, denotes the indicator function; Step S4 specifically includes: the purpose of using data enhancement is to further adjust the distribution of the feature space using new data; the data enhancer is represented as The enhanced data is represented as The enhanced sample is obtained by random discarding, and the idea is to copy the original sample first, as follows: (5) Then, the discarded point number is set as ; in , randomly select non-repeating values, and form an index sequence ; then update as follows: (6) The newly generated sample is used to update the model after similarity screening. The specific process is as follows: first, encode the negative enhancement sample using the encoder and calculate the distance vector: (7) Then, normalize the distance vector using the softmin function, and then screen the samples that meet the following rules: (8) wherein, is the class of the sample corresponding to denotes a slicing operation, i.e. taking the first elements and constructing a new vector; and is a hyper-parameter controlling the similarity; the purpose of using the slicing operation is to mask the pseudo prototype points, in order to calculate whether the enhanced sample's feature vector keeps a certain similarity with its corresponding class prototype; Step S5 specifically includes: for the samples from the known class, the preliminary loss function is calculated as follows: (9) For the enhanced samples retained by the similarity screening, the preliminary loss function is calculated as follows: (10) In addition, in order to minimize the inter-class distance, a distance loss function restriction term is added as follows: (11) Finally, the loss of each sample is calculated as follows: (12) wherein and is an adjustable hyperparameter, setting both parameters to 0.

1.

2. The method of claim 1, wherein the method is based on data-negative enhancement and pseudo-class prototype points for open-set identification of electromagnetic signals. Step S2 specifically comprises: in order to improve the robustness of the distribution of the feature space, the position of the class prototype of each class is determined at the initialization time and remains unchanged in the training stage; the class prototype point constructed in this way is called a class anchor point; using this idea, a class prototype point is preset for each known class; in order to increase the ability of the model to cope with the risk of an open world, a new prototype point is additionally added, which is called a pseudo-class prototype point; thus, the prototype point of the i-th class is represented as k the prototype point of the i-th class is represented as the prototype point of the i-th class is represented as the prototype point of the i-th class is represented as (2) wherein is a hyper-parameter that determines the location of the initial class prototype points; during the training process, all class prototype points are not updated, but after the training is completed, the class prototype points of the known classes are updated again using the training samples, and the values of the pseudo-class prototype points are kept unchanged.

3. The method of claim 1, wherein the method is based on data-negative enhancement and pseudo-class prototype points for open-set identification of electromagnetic signals. Step S3 specifically comprises: representing the neural network encoder of the whole model as , wherein represents a feature space of the open set identification model, is a dimension of the feature space; a feature vector of a sample in the feature space is obtained by inputting the sample into ; a Euclidean distance is used as a distance measurement manner, and a distance calculation function is represented as ; then a distance vector composed of distances between the feature vector and each class anchor point is calculated, wherein: (3) Then, normalize the distance vector using the softmin function as follows: (4)。 4. The method of claim 1, wherein the method is based on data-negative enhancement and pseudo-class prototype points for open-set identification of electromagnetic signals. Step S6 specifically includes: First, after training, for each known class, the average of the feature vectors of all training samples of that known class is used as the class prototype; second, all correctly identified known class samples are selected... The average value is used as the base threshold. ​ Ultimately, during testing, for a single test sample... The distance vector is obtained by encoding through an encoder. Then, the sample is judged using the following judgment method: (13)。

Images