Electromagnetic signal classification method based on case-based selection and multi-objective linear programming
By constructing an incremental learning electromagnetic signal classification model based on class paradigm selection and multi-objective linear programming, this paper solves the problem that existing models cannot learn new categories, achieves efficient classification of old and new categories, and quickly updates the electromagnetic signal classifier.
Patent Information
- Application Number
- CN202210317604.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-29
- Publication Date
- 2026-01-02
- Estimated Expiration
- 2042-03-29
AI Technical Summary
Existing electromagnetic signal classification models cannot incrementally learn new signal categories, especially when faced with open datasets, and struggle to maintain their classification ability for older categories.
An electromagnetic signal classification model based on class paradigm selection and multi-objective linear programming is constructed by reordering and filtering training samples. The classifier is updated using a deep convolutional neural network and multi-objective linear programming, and feature fine-tuning training is performed by combining a weighted loss function of distillation loss and cross-entropy loss.
It achieves high classification accuracy for new categories while maintaining the ability to classify old categories, quickly learns new category data, reduces the number of iterations for training, and improves the learning speed.
Smart Images
Figure CN116933050B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of signal processing, in particular to an electromagnetic signal classification method based on example selection and multi-objective linear programming. BACKGROUND
[0002] As a key technology in the field of information processing, electromagnetic signal classification is the basis of non-cooperative communication, intelligent antenna and software radio. Electromagnetic signal classification includes identifying electromagnetic signal types and identifying modulation information. With the widespread use of various radio devices, the types and number of radiation sources in the electromagnetic environment are increasing, and the electromagnetic spectrum is becoming more crowded. More importantly, the speed of capturing electromagnetic signal data is faster than before. In this case, electromagnetic signal data is widely used in electromagnetic spectrum detection, cognitive radio, network space security and many other fields, and electromagnetic signal classification is an important prerequisite, so electromagnetic signal classification has very important research value. With the increase in the number of signal processing tasks, electromagnetic signal data is showing an explosive growth trend. Especially in the face of new types of electromagnetic signal data. In recent years, with the rapid development of deep learning, many related algorithms have been used for electromagnetic signal classification, but most of the existing algorithms are for closed data sets, and cannot identify new signal classes. In real applications, more open data sets are faced, and the classification model must be able to quickly learn new class data while maintaining the classification ability of old class data. SUMMARY
[0003] In view of the above analysis, the present application aims to provide a method for predicting the communication quality of a UAV link, which solves the problem that existing electromagnetic signal classification models cannot incrementally learn new signal classes.
[0004] The technical scheme provided by the present application is:
[0005] The present application discloses an electromagnetic signal classification method based on example selection and multi-objective linear programming, comprising:
[0006] Old class data of electromagnetic signals including multiple classes is used to train a first classification model; the first classification model is a deep convolutional neural network classification model;
[0007] The training samples of the old class data are reordered according to the normalized mutual information between the real labels and the predicted labels in the training; and according to the classification difficulty, part of the training samples in the old class data of each class are sequentially screened and stored;
[0008] For new class data of an added class, multi-objective linear programming is performed on the basis of the first classification model, the classifier of the first classification model is updated, and a second classification model is obtained;
[0009] The feature fine-tuning training of the second classification model is performed on the training samples of the stored old class data and the added new class data, and an incremental learning classification model is obtained, which is used for classifying the electromagnetic signal data of the new class and / or the old class.
[0010] Further, in the first classification model training process, the Ground truth corresponding to the training sample set of the old class data is used as the label of the training sample set, and the gradient descent method is used for iterative supervised training of the deep convolutional neural network model to obtain the trained first classification model.
[0011] Further, the old class data training samples are reordered according to the normalized mutual information between the real labels and the predicted labels in the training, including:
[0012] The normalized mutual information between the one-hot label corresponding to the Ground truth of the training set of the old class data and the predicted probability vector is calculated Where I(y,y) is the mutual information between the one-hot label corresponding to the Ground truth and the predicted probability vector y; H(y) is the entropy of the real label; H(y) is the entropy of the predicted probability;
[0013] All training samples of each class in the old class data training set are sorted in ascending order according to the normalized mutual information.
[0014] Further, part of the training samples are filtered and stored in order according to the classification difficulty in the old class data of each class, including:
[0015] 1) The classification difficulty measurement factor of each class is obtained by predicting all training samples in the old class data using the trained first classification model;
[0016] 2) The sample filtering number of each class is determined according to the size of the memory space for storing training samples combined with the classification difficulty measurement factor of each class;
[0017] 3) According to the sample filtering number of each class, the corresponding number of samples with small normalized mutual information are filtered and stored from the sorted training data of each class.
[0018] Further, the calculation method of the classification difficulty measurement factor of each class includes:
[0019] The features of all training data are extracted using the trained first classification model, and the average feature vector of each class is calculated;
[0020] The predicted classification accuracy and the average distance of each training sample to the center of its own class are used as the classification difficulty measurement factor of each class;
[0021] The classification difficulty measurement factor is scaled to 0 to 1 by using a softmax function.
[0022] Further, the classification difficulty measurement factor of each class is determined by the following formula: wherein, η i is the classification difficulty measurement factor of the i-th class, OA i is the classification accuracy of the i-th class; n i is the number of training samples of the i-th class; f k is the feature vector of the training sample, is the average feature vector of the i-th class sample.
[0023] Further, the number of samples to be screened for each class is determined by combining the memory space size for storing the training samples with the classification difficulty measurement factor of each class wherein m i represents the number of samples of the i-th class to be screened, and K is the memory space size for storing the training samples.
[0024] Further, the weight matrix of the classifier of the first classification model is:
[0025]
[0026] The weight matrix of the classifier of the second classification model after adding T weight columns in the classifier of the first classification model is:
[0027]
[0028] wherein, W represents the weight, N is the number of old classes, T is the number of new classes, l is the serial number of the feature vector, and L is the dimension of the feature vector.
[0029] Further, a multi-objective linear programming problem is constructed according to the classification principle of the linear classifier, and the weight matrix W of the classifier of the second classification model is obtained by solving the multi-objective linear programming problem new ;
[0030] The multi-objective linear programming problem is:
[0031]
[0032] wherein, f j is the average feature vector of the j-th new class, f i is the average feature vector of the i-th old class, W j represents the j-th column of the new weight matrix, and W k represents the j-th column of the new weight matrix; max{|W old |} represents the maximum value of the old weight matrix.
[0033] Further, in the feature fine-tuning training of the second classification model using the stored old class data and new class data training samples, the weighted cross-entropy loss and distillation loss are used as the overall loss function.
[0034] Advantages of the present application:
[0035] The electromagnetic signal classification model constructed by the present application can incrementally learn new electromagnetic signal classes, not only achieving high classification accuracy for new classes, but also maintaining the classification ability for old electromagnetic signal classes, overcoming the problem that existing electromagnetic signal classification models cannot learn new classes.
[0036] The present application only uses the class average feature vector extracted and calculated by the forward calculation of the convolutional neural network to construct and solve the multi-objective linear programming problem to obtain the incremental classifier, without the need for a large number of iterative training, and only a small number of fine-tuning training is required to quickly learn new class data, speeding up the learning speed of new classes.
[0037] The present application uses the normalized mutual information between the real labels of the output data of the model as the selection standard, and selects different numbers of samples according to the different classification difficulties of different classes; the sample selection method is used in cooperation with the loss function based on the weighted distillation loss and cross-entropy loss proposed by the present application to effectively ensure that the classification ability of the model for old classes decreases less. BRIEF DESCRIPTION OF DRAWINGS
[0038] The accompanying drawings are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiments of the application and together with the description serve to explain the principles of the application.
[0039] Figure 1 The electromagnetic signal classification method based on class example selection and multi-objective linear programming in the embodiment of the present application is shown in the flowchart.
[0040] Figure 2 The process diagram of obtaining a new classifier using multi-objective linear programming in the embodiment of the present application is shown. DETAILED DESCRIPTION
[0041] The preferred embodiments of the present application will be specifically described below in conjunction with the accompanying drawings, wherein the drawings constitute a part of this application and are used together with the embodiments of the present application to explain the principles of the present application.
[0042] The electromagnetic signal classification method based on class example selection and multi-objective linear programming disclosed in the present embodiment is shown in the flowchart, which includes: Figure 1
[0043] Step S1, using old class data of electromagnetic signals including multiple classes, training a first classification model; the first classification model is a deep convolutional neural network classification model;
[0044] Step S2, reordering the training samples of the old class data according to the normalized mutual information between the real label and the predicted label in the training; and storing part of the training samples in order in the old class data of each class according to the classification difficulty;
[0045] Step S3, for the new class data of the increased class, updating the classifier of the first classification model by multi-objective linear programming based on the first classification model to obtain a second classification model;
[0046] Step S4, performing feature fine-tuning training of the second classification model using the stored training samples of the old class data and the increased new class data to obtain an incremental learning classification model for classifying the electromagnetic signal data of the new class and / or the old class.
[0047] Specifically, in step S1, the old class data of the electromagnetic signal including N classes is used to construct a training set and a test set of the old class data.
[0048] More specifically, an N-class electromagnetic signal sample set and the class to which each electromagnetic signal sample belongs are obtained, and the electromagnetic signal sample set is divided into a training set and a test set, each accounting for 50%, wherein the dimension of each sample is 4096x2.
[0049] In step S1, the first classification model constructed is a deep convolutional neural network classification model including a plurality of convolutional neural networks, a fully connected network, a Dropout network and an output layer network connected in sequence.
[0050] Specifically, the deep convolutional neural network classification model includes seven convolutional neural networks; each convolutional neural network includes a convolutional layer, a regularization layer and a pooling layer.
[0051] More specifically, the embodiment constructs a deep convolutional neural network classification model containing 24 layers: first convolutional layer→first regularization layer→first pooling layer→second convolutional layer→second regularization layer→second pooling layer→third convolutional layer→third regularization layer→third pooling layer→fourth convolutional layer→fourth regularization layer→fourth pooling layer→fifth convolutional layer→fifth regularization layer→fifth pooling layer→sixth convolutional layer→sixth regularization layer→sixth pooling layer→seventh convolutional layer→seventh regularization layer→seventh pooling layer→first fully connected layer→first Dropout layer→output layer. According to the input data size requirement, the designed deep fully convolutional neural network model contains 24 layers.
[0052] Specifically, the parameter settings of each layer are as follows: the step length of the first convolutional layer is set to 1, and the step lengths of the second to seventh convolutional layers are all set to 2; the kernel size of the first convolutional layer is set to 15x2, the kernel size of the second convolutional layer is set to 7x1, the kernel sizes of the third, fourth, sixth and seventh convolutional layers are all set to 5x1, the kernel size of the fifth convolutional layer is set to 3x1, the number of all kernels is 150, all the pooling layers adopt the maximum pooling manner, the size of the pooling kernel is set to 2x1, and the step length of the pooling layer is set to 2; the number of output neurons of the first fully connected layer is 512; the ratio of the first Dropout layer is set to 0.5; and the number of output neurons of the output layer is N. That is, the classifier of the deep fully convolutional neural network model can classify the old class data of N classes.
[0053] In the first classification model training process, the constructed training set is used as the input of the first classification model, the Ground truth corresponding to the training sample set of the old class data is used as the label of the training sample set, the gradient descent method is used for iterative supervised training of the model, and a trained deep convolutional neural network model is obtained. The first classification model can achieve high classification accuracy on N old data.
[0054] In step S2, the constructed training set of N classes of old class data is input into the first classification model to obtain the prediction probability vector of all data; the training samples of the old class data are reordered according to the normalized mutual information between the real label and the predicted label in the training, including:
[0055] 1) Calculate the normalized mutual information between the one-hot label corresponding to the Ground truth and the prediction probability vector of the training set of the old class data Where I(y,y) is the mutual information between the one-hot label corresponding to the Ground truth and the prediction probability vector y; H(y) is the entropy of the real label; and H(y) is the entropy of the prediction probability.
[0056] Mutual information between one-hot label corresponding to Ground truth and prediction probability vector y
[0057] Entropy of real label
[0058] Entropy of prediction probability
[0059] p(y) is the probability of the real label. is the prediction probability.
[0060] 2) Sort all training samples of each class in the old class data training set in ascending order of normalized mutual information.
[0061] Specifically, in step S2, according to the classification difficulty, part of the training samples in the old class data of each class are stored, including:
[0062] 1) Using the trained first classification model to predict the training samples of all classes in the old class data to obtain the classification difficulty measurement factor of each class;
[0063] Wherein, the process of obtaining the classification difficulty measurement factor of each class includes:
[0064] (1) Using the trained first classification model to extract the features of all training data, and calculating the average feature vector of each class;
[0065] (2) Using the predicted classification accuracy and the average distance of each training sample to the center of its own class as the classification difficulty measurement factor of each class;
[0066] Classification difficulty measurement factor Wherein, η i is the classification difficulty measurement factor of the i-th class, OA i is the classification accuracy of the i-th class; n i is the number of training samples of the i-th class; f k is the feature vector of the training sample, is the average feature vector of the i-th class sample.
[0067] (3) Using the softmax function to scale the classification difficulty measurement factor to 0 to 1.
[0068] 2) According to the memory space size used to store the training samples combined with the classification difficulty measurement factor of each class to determine the sample screening number of each class;
[0069] Specifically, the sample screening number of each class is determined Wherein, m i represents the number of samples to be screened for the i-th class, and K is the memory space size for storing training samples.
[0070] 3) According to the sample screening number of each class, the corresponding number of samples with small normalized mutual information are screened from the sorted training data of each class and stored.
[0071] Specifically, in step S3, the new class data of T classes of electromagnetic signals not contained in the old class data is obtained; and the training set and test set of the new class data are constructed.
[0072] For the new class data of T classes of electromagnetic signals obtained and the class to which each electromagnetic signal sample belongs, it is divided into a training set and a test set, wherein the training set and the test set each account for 50%.
[0073] For the new class data of the increased multiple classes, on the basis of the first classification model, multi-objective linear programming is performed, the feature extractor in the first classification model is reserved, the old classifier is incremented, and a second classification model including an incremental classifier is obtained.
[0074] As shown in the process diagram for obtaining a new classifier by using multi-objective linear programming, in the process of obtaining the new classifier, the feature vectors of the new class training data and the saved old class data are extracted through forward calculation of the trained classification model, and the class average feature vectors of all classes are calculated. Figure 2
[0075] In order to enable the classifier to recognize the new class, the number of output nodes of the output layer in the second classification model is increased to N+T, that is, T weight columns are added to the weight matrix of the original classifier.
[0076] The weight matrix of the classifier of the first classification model:
[0077]
[0078] The weight matrix of the classifier of the second classification model after T weight columns are added to the classifier of the first classification model:
[0079]
[0080] Wherein, W represents the weight, N is the number of old classes, T is the number of new classes, l is the serial number of the feature vector, and L is the dimension of the feature vector.
[0081] Then, a multi-objective linear programming problem is constructed according to the classification principle of the linear classifier, and the weight matrix of the new classification layer is obtained by solving the problem, so as to obtain a new second classification model.
[0082] The multi-objective linear programming problem is:
[0083]
[0084] Wherein, f j is the average feature vector of the jth new class, f i is the average feature vector of the ith old class, W j represents the jth column of the new weight matrix, and W k represents the jth column of the new weight matrix; max{|W old |} represents the maximum value of the old weight matrix.
[0085] In step S4, the stored old class data and new class data training samples are used for feature fine-tuning training of the second classification model, so as to obtain an incremental learning classification model, which is used for classifying new class data and old class data.
[0086] During the feature fine-tuning training process, the cross-entropy loss and the weighted distillation loss are taken as the overall loss function.
[0087] After the feature fine-tuning training of the second classification model, classification tests are performed on the test sets of the new class and the old class.
[0088] The old class test samples and the new class test samples are respectively input into the new classification model, and the output layer outputs the probability of each class to which the sample belongs through the softmax activation function, and the class corresponding to the maximum probability value is the class predicted by the model.
[0089] To sum up, the electromagnetic signal classification model constructed by the embodiment of the application can incrementally learn new electromagnetic signal classes, not only achieving high classification accuracy for new classes, but also maintaining the classification ability for old electromagnetic signal classes, overcoming the problem that existing electromagnetic signal classification models cannot learn new classes. Only the class average feature vector extracted and calculated by the forward calculation of the convolutional neural network is used to construct and solve the multi-objective linear programming problem to obtain the incremental classifier, without the need for a large number of iterative training, and only a small number of fine-tuning training batches are needed to quickly learn new class data, accelerating the learning speed of new classes. The normalized mutual information between the real labels of the output data of the model is used as the selection standard, and different numbers of samples are selected according to the different classification difficulties of different classes; the sample screening method is used in cooperation with the loss function based on the distillation loss and the cross-entropy loss weighting proposed by the application, which can effectively ensure that the classification ability of the model for the old class decreases less.
[0090] The above is only a preferred specific embodiment of the application, but the protection scope of the application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the application, which should be covered within the protection scope of the application.
Claims
1. An electromagnetic signal classification method based on class paradigm selection and multi-objective linear programming, characterized in that, include: The first classification model is trained using old class data of electromagnetic signals that include multiple classes. The first classification model is a deep convolutional neural network classification model; The training samples of the old class data are reordered based on the normalized mutual information between the true labels and the predicted labels during training. Based on the classification difficulty, a portion of the training samples are selected and stored sequentially from the old class data of each class. For new class data with added classes, multi-objective linear programming is performed based on the first classification model to update the classifier of the first classification model and obtain the second classification model; The second classification model is fine-tuned using training samples of stored old class data and added new class data to obtain an incrementally learned classification model, which is used to classify electromagnetic signal data of new and / or old classes. Based on the classification difficulty, a subset of training samples are sequentially selected from the old class data for each class and stored, including: 1) Use the trained first classification model to predict the training samples of all classes in the old class data to obtain the classification difficulty measure factor for each class; Classification difficulty measurement factor Where, η i Let OA be the classification difficulty measure for the i-th class. i Let n be the classification accuracy for the i-th class; i f is the number of training samples for the i-th class of data; k For the feature vectors of the training samples, Let be the average feature vector of the i-th class of samples; 2) Determine the number of samples to be selected for each class based on the memory space used to store training samples and the classification difficulty measure factor for each class; Where m i This represents the number of samples to be filtered in the i-th class, and K is the memory space size for storing the training samples; 3) Based on the number of samples to be selected for each class, select and store the corresponding number of samples with smaller normalized mutual information from the training data of each class; Based on the classification principle of linear classifiers, a multi-objective linear programming problem is constructed and solved to obtain the weight matrix W of the second classification model classifier. new ; The multi-objective linear programming problem is: Among them, f j f is the average eigenvector of the j-th new class. i W is the average eigenvector of the i-th old class. j Represents the new weight matrix and W in the j-th column. k represents column j of the new weight matrix; max{|W old |} represents the maximum value of the old weight matrix.
2. The electromagnetic signal classification method according to claim 1, characterized in that, During the training of the first classification model, the ground truth corresponding to the training sample set of the old class data is used as the label of the training sample set. Gradient descent is used to perform iterative supervised training on the deep convolutional neural network model to obtain the trained first classification model.
3. The electromagnetic signal classification method according to claim 1, characterized in that, The training samples of the old class data are reordered based on the normalized mutual information between the true and predicted labels during training, including: Calculate the normalized mutual information between the one-hot labels and the predicted probability vectors corresponding to the ground truth of the training set of the old class data. in, The one-hot label and predicted probability vector corresponding to the ground truth Mutual information; The entropy of the real label; Entropy for predicting probabilities; Sort all training samples of each class in the old class training set according to their normalized mutual information from smallest to largest.
4. The electromagnetic signal classification method according to claim 1, characterized in that, The calculation methods for the classification difficulty factor for each class include: The trained first classification model is used to extract features from all training data, and the average feature vector of each class is calculated. The predicted classification accuracy and the average distance from each training sample to its own class center are used as the classification difficulty measure factors for each class. The softmax function is used to scale the classification difficulty measure factor to between 0 and 1.
5. The electromagnetic signal classification method according to claim 1, characterized in that, Weight matrix of the classifier in the first classification model The weight matrix of the second classification model classifier is obtained by adding T weight columns to the classifier of the first classification model: Where W represents the weight, N is the number of old classes, T is the number of new classes, l is the feature vector index, and L is the dimension of the feature vector.
6. The electromagnetic signal classification method according to claim 1, characterized in that, In the feature fine-tuning training of the second classification model using training samples of stored old and new class data, the weighted sum of cross-entropy loss and distillation loss is used as the overall loss function.