Method for intelligent identification of electromagnetic spectrum radiation source modulation mode based on unbalanced learning

Abstract

The application discloses a kind of electromagnetic spectrum radiation source modulation mode intelligent identification method based on unbalanced learning, including the training phase and inference phase of model, first establish database, then modulated signal sequence is converted into the form of IQ signal, normalization is carried out, data label is converted into the form of one-hot encoding label;According to the soft confusion matrix calculated by verification set data;Based on soft confusion matrix and training set one-hot label, training set enhanced coding label is obtained;Training set data is input into model, and weight is updated according to gradient descent method;Training is finished, and electromagnetic spectrum radiation source modulation mode intelligent identification model with self-balancing characteristics is obtained;When data forward inference is carried out, the received signal is preprocessed and input into the model, and the modulation mode of the radiation source is automatically determined according to the model output.The method disclosed in the application can effectively improve the radiation source modulation mode identification effect of the model under the condition of unbalanced training data.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent identification of modulation modes of radiation sources, and in particular to an electromagnetic spectrum radiation source modulation mode intelligent identification method based on unbalanced learning. BACKGROUND

[0002] With the development and application of electronic countermeasure related technologies, the demand for identifying radiation source signals such as communication, navigation, radar, remote sensing and telemetry, and jamming is growing. Traditional radiation source identification technology mainly relies on manual feature analysis, which consumes a lot of manpower and resources. In order to improve work efficiency and identification effect, new radiation source modulation identification methods based on machine learning and artificial intelligence technology have gradually become a research hotspot and achieved fruitful results. The effect of machine learning algorithm depends on the quantity and quality of model training data, and the unbalanced problem that may exist in the data set will also deeply affect the effect of spatial electromagnetic spectrum radiation source modulation mode intelligent identification method.

[0003] In view of the data imbalance problem, many unbalanced algorithms have been proposed, which are mainly divided into two categories: data level method and algorithm level method. The data level method mainly adjusts the distribution of data to make the originally unbalanced training data balanced, and the algorithm level method mainly changes the algorithm itself to make the model achieve good recognition effect even in the case of unbalanced training data. There is no research on applying unbalanced algorithms to spatial electromagnetic spectrum radiation source modulation mode intelligent identification. SUMMARY

[0004] To solve the above technical problems, the present application provides a spatial electromagnetic spectrum radiation source modulation mode intelligent identification method based on unbalanced learning, which belongs to an algorithm level unbalanced learning method, to effectively improve the radiation source modulation mode identification effect of the model under the condition of unbalanced training data.

[0005] To achieve the above purpose, the technical scheme of the present application is as follows:

[0006] A spatial electromagnetic spectrum radiation source modulation mode intelligent identification method based on unbalanced learning, comprising the following processes: a training phase and an inference phase of a spatial electromagnetic spectrum radiation source modulation mode intelligent identification model, wherein the intelligent identification model adopts a deep neural network model;

[0007] The training phase of the model comprises the following steps:

[0008] (1) Establish a wide-area spatial electromagnetic spectrum database including electromagnetic signal data to generate a data set required for model training;

[0009] (2) For the input modulation signal, convert it into the form of IQ signal and generate the data set required for training the model in the form of a dual-channel sequence;

[0010] (3) Normalize the input data feature values ​​and convert all data labels into one-hot encoded labels. Then divide the dataset into three sets: training set, validation set, and test set.

[0011] (4) Before each training round of the model, the validation set data is input into the model. The soft confusion matrix is ​​calculated based on the model output and the one-hot encoded labels of the validation set, and the imbalance features of the data and the model are extracted.

[0012] (5) Based on the soft confusion matrix, update the enhanced coding generation matrix of the training set, and perform matrix multiplication between the enhanced coding generation matrix and the one-hot coding label of the training set to obtain the enhanced coding label of the training set.

[0013] (6) Input the training set data into the model, calculate the loss function based on the model output and the training set augmented coding labels obtained in the previous steps, then perform backpropagation of the error based on the loss function, find the gradient of the loss function on each weight of the model, and update the weights according to the gradient descent method.

[0014] (7) Repeat steps (4) to (6) in each training round until the training ends. Use the test set data to evaluate the model and finally obtain the intelligent recognition model of electromagnetic spectrum radiation source modulation mode.

[0015] The inference phase of the model includes the following process:

[0016] The system receives electromagnetic signal data, performs IQ conversion and normalization preprocessing on the electromagnetic signal data, and then inputs it into the trained model. Based on the model output, it automatically determines the modulation mode of the radiation source.

[0017] In the above scheme, in step (2), the modulation signal is converted into the form of IQ signal, where I is the in-phase component and Q is the quadrature component.

[0018] In a further technical solution, the mathematical relationship between the modulation signal and the IQ signal is as follows:

[0019] x(t)=I(t)cos(2πf c t)+Q(t)sin(2πf c t)

[0020] Where x(t) is the modulation signal, I(t) and Q(t) are the I-channel signal and Q-channel signal, respectively, and f c Let t be the carrier frequency and t be the time.

[0021] In the above scheme, in step (3), the input data feature values are normalized for pretreatment, and the value range is scaled to the interval [-1, 1]; for the data label, it is converted into a one-hot encoding form, and if a certain example belongs to the p-th class, its one-hot encoding label is described as:

[0022]

[0023] wherein, l o is a one-hot encoding label, N is the total number of classes in the data set, The value is as follows:

[0024]

[0025] wherein, p is the index of the class to which the example belongs, represents the q-th element of l o .

[0026] In the above scheme, in step (4), the calculation method of the soft confusion matrix S is as follows:

[0027] S=L oT Y

[0028] wherein, L o is a matrix composed of one-hot encoding labels of all examples in the set, Y is a matrix composed of model output values based on the examples in the set, and the values of their elements are as follows:

[0029]

[0030] y i,q =P(q|x i )

[0031] The elements of the soft confusion matrix S are defined as follows:

[0032]

[0033] wherein, P(q|x i ) represents the probability that the example i predicted by the model belongs to the q-th class, D p is a set of indexes of all examples belonging to the p-th class, that is, i∈D p represents that the example i belongs to the p-th class, and s p,q represents the element in the p-th row and the q-th column of the soft confusion matrix S.

[0034] In the above scheme, in step (5), the enhanced encoding generation matrix is represented as follows:

[0035] G=(1-μ)G * +μ(I-∈C)

[0036] where G * is the old value of the enhanced encoding generation matrix in a training round, which is initialized as an N-by-N identity matrix in the first training round; μ and ∈ are the update rate and the enhancement rate, respectively; I is an N-by-N identity matrix; N is the total number of classes in the dataset; and C is a cost matrix, which is obtained based on a soft confusion matrix of a validation set:

[0037] C=S'-I

[0038] where S' is a row-normalized soft confusion matrix, and an element of S' is:

[0039]

[0040] training set enhanced encoding label is calculated as follows:

[0041]

[0042] where is a matrix composed of one-hot encoding labels of all samples in the training set, and the matrix multiplication is defined based on a real number field.

[0043] In the above scheme, in step (6), the cross-entropy loss function calculated based on the enhanced encoding label is as follows:

[0044]

[0045] where is an element of the training set enhanced encoding label , N is the total number of classes in the dataset, y i,q is an element of the model output vector, and q is the serial number of the class.

[0046] In the above scheme, in the model inference stage, the output of the model is a prediction vector, and each element of the vector represents the prediction that the input signal modulation mode belongs to the class, and the prediction vector is:

[0047] y=[y0,y1,y2,…,y q ,…,y N-1 ]

[0048] y q =P(q|x)

[0049] where x represents the model input, y represents the model output, P(q|x) represents the model prediction probability that the input signal modulation mode belongs to the class q, and the modulation decision follows the criterion that the prediction probability is the largest:

[0050]

[0051] Wherein, p is the modulation mode category of the input radiation source signal finally judged by the model.

[0052] Through the technical scheme, the electromagnetic spectrum radiation source modulation mode intelligent identification method based on unbalanced learning has the following beneficial effects:

[0053] (1) The input sequence feature is converted into the form of an IQ signal, and normalized, so that the model can extract effective deep features therefrom;

[0054] (2) The present application proposes an improved version of the soft confusion matrix to address the pain point of requiring a large amount of data for the confusion matrix. Unlike the former, which counts the number of samples belonging to different cases, the latter sums the prediction probabilities of the samples belonging to different cases. This makes the soft confusion matrix also achieve good results in the case of less data, while also reducing the computational cost.

[0055] (3) The present application discards the practice of using one-hot encoding labels for data labeling and calculating loss functions during training, and instead proposes to replace them with enhanced encoding labels for error backpropagation and parameter updating;

[0056] (4) When generating training data labels, the present application uses matrix multiplication of a generation matrix and a real number domain matrix to enhance the encoding of the labels. For the generation matrix update of each training round, the present application retains a portion of the old value and updates a portion of the new value based on the soft confusion matrix calculation, in order to reduce the instability generated during training.

[0057] (5) The soft confusion matrix used by the present application can reflect the fitting degree of the model to different categories of data and the proportion of errors in different cases to some extent. Therefore, the enhanced encoding labels generated based on the soft confusion matrix can enhance the fitting degree of the model to under-fitting category data and reduce major errors, thereby training a spatial electromagnetic spectrum radiation source modulation mode intelligent identification model with high generalization ability. BRIEF DESCRIPTION OF DRAWINGS

[0058] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows.

[0059] Figure 1 A schematic diagram of the electromagnetic spectrum radiation source modulation mode intelligent identification method based on unbalanced learning disclosed by the present application;

[0060] Figure 2 IQ signals of some common modulation modes;

[0061] Figure 3A way for generating a modulated signal through an IQ signal;

[0062] Figure 4 A flowchart for calculating a soft confusion matrix;

[0063] Figure 5 A flowchart for label enhancement coding of a training set. DETAILED DESCRIPTION

[0064] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application.

[0065] The present application provides an electromagnetic spectrum radiation source modulation mode intelligent identification method based on unbalanced learning, as shown in the formula: Figure 1 The present application provides an electromagnetic spectrum radiation source modulation mode intelligent identification method based on unbalanced learning, as shown in the formula:

[0066] I. Training phase of the model

[0067] 1. Database establishment

[0068] According to the characteristics of electromagnetic threats under the electronic warfare situation, a wide-area spatial electromagnetic spectrum database including communication, navigation, radar, remote sensing and telemetry, jamming and other electromagnetic signal data is established, and a data set required for model training is generated in combination with big data related technologies.

[0069] 2. Data preprocessing

[0070] For the input modulated signal, it is converted into the form of an IQ signal, and a data set required for training the model is generated in the form of a double-channel sequence.

[0071] In the IQ signal, I is the in-phase component, which can be regarded as the projection of the modulated signal on cos(2pft) (f is the carrier frequency), and Q is the quadrature component, which can be regarded as the projection of the modulated signal on -sin(2pft). c c c Figure 2 IQ signals of some common modulation modes are shown.

[0072] A schematic diagram for generating a modulated signal through an IQ signal is shown in the formula: Figure 3 The mathematical relationship is as follows:

[0073] x(t) = I(t)cos(2pft) + Q(t)sin(2pft) c c

[0074] ​​​​​where x(t) is the modulated signal, I(t) and Q(t) are the I and Q signals respectively, f c is the carrier frequency, and t is time.

[0075] For the needs of the deep neural network model training process, the input data feature values are normalized and preprocessed, and the value range is scaled to the interval [-1, 1]; and all data labels are converted into one-hot encoding label form, if a sample belongs to the p-th class, its one-hot encoding label is described as:

[0076]

[0077] where l o is the one-hot encoding label, N is the total number of classes in the data set, The values are as follows:

[0078]

[0079] where p is the index of the class to which the sample belongs, represents the q-th element of l o .

[0080] Then the data set is divided into training set, validation set and test set.

[0081] 3. Calculate the soft confusion matrix

[0082] Before each training round of the model, the validation set data is input into the model, and the soft confusion matrix is calculated according to the model output and the validation set one-hot encoding label, and the unbalanced features of the data and the model are extracted.

[0083] The process of calculating the soft confusion matrix is shown in Figure 4 The calculation method of the soft confusion matrix S is as follows:

[0084] S = L oT Y

[0085] where L o is a matrix composed of one-hot encoding labels of all samples in the set, Y is a matrix composed of model output values based on the samples in the set, and their element values are as follows:

[0086]

[0087] y i,q = P(q|x i )

[0088] The elements of the soft confusion matrix S are defined as follows:

[0089]

[0090] where P(q | x i ) denotes the probability that the instance i belongs to class q predicted by the model, D p is the index set of all instances belonging to class p, i.e. i ∈ D p denotes that instance i belongs to class p, s p,q denotes the element of the p-th row and q-th column of the soft confusion matrix S.

[0091] In order to reduce the adverse effects of the imbalance of the number of instances of different classes, the soft confusion matrix needs to be row-normalized, and the row-normalized element of the soft confusion matrix is expressed as follows:

[0092]

[0093] where s′ p,q denotes the element of the row-normalized soft confusion matrix S′, s p,q denotes the element of the soft confusion matrix before row normalization.

[0094] 4. Enhanced encoding

[0095] The enhanced encoding process of the training set label is shown in Figure 5 , which includes the following processes:

[0096] (1) Based on the normalized soft confusion matrix, the cost matrix C is calculated:

[0097] C = S′ - I

[0098] where S′ is the row-normalized soft confusion matrix based on the validation set, I is an N-row and N-column unit matrix, and N is the total number of classes in the data set.

[0099] (2) Update the enhanced encoding generation matrix:

[0100] G = (1 - μ)G * + μ(I - ∈C)

[0101] where G * is the old value of the enhanced encoding generation matrix in the last training round, and is initialized as an N-row and N-column unit matrix in the first training round; the hyperparameters μ and ∈ are the update rate and the enhancement rate, respectively.

[0102] (3) Perform matrix multiplication between the enhanced encoding generation matrix and the one-hot encoding label of the training set to obtain the enhanced encoding label of the training set

[0103]

[0104] where, is a matrix composed of one-hot encoding labels of all samples in the training set, and the matrix multiplication is defined based on the real number field.

[0105] 5. Model training and parameter updating

[0106] The training set data is input into the model, the loss function is calculated according to the model output and the enhanced encoding label of the training set obtained in the previous step, and then error back propagation is performed based on the loss function, the gradient of the loss function on each weight of the model is calculated, and the weight is updated according to the gradient descent method.

[0107] Wherein, the cross-entropy loss function calculated based on the enhanced encoding label is as follows:

[0108]

[0109] Wherein, is an element of the enhanced encoding label of the training set , N is the total number of categories in the data set, y i,q is an element of the model output vector, and q is the serial number of the category.

[0110] The partial derivative of y i,q is:

[0111]

[0112] It can be seen that the value of can affect the partial derivative of the loss function, and according to the error back propagation algorithm, this value will also affect the update of the model weight.

[0113] In each training round, the process of soft confusion matrix calculation, training set label enhanced encoding update, model weight parameter update and the like is repeated until the training is completed, and then the model is evaluated using the test set data, and finally an electromagnetic spectrum radiation source modulation mode intelligent recognition model with high generalization ability is obtained.

[0114] The model obtained by training has a self-balancing feature, that is, even if the training data is unbalanced, the model can achieve a relatively balanced recognition effect, that is, the recognition performance of each type of radiation source modulation mode is similar.

[0115] II. Inference stage of the model

[0116] The obtained model is used for radiation source modulation mode recognition. When performing data forward inference, electromagnetic signal data is received, the electromagnetic signal data is subjected to IQ conversion and normalization preprocessing, and then it is input into the model obtained after training. According to the model output, the modulation mode of the radiation source is automatically determined.

[0117] The output of the model is a prediction vector, and each element of the vector represents the prediction that the input signal modulation belongs to the class:

[0118] y = [y0, y1, y2,..., y q ,..., y N-1 ]

[0119] y q = P(q | x)

[0120] Wherein, x represents the input of the model, y represents the output of the model, P(q | x) represents the prediction probability of the input signal modulation belonging to the class q, and the modulation decision follows the criterion of the maximum prediction probability:

[0121]

[0122] Wherein, p is the modulation class of the input radiation source signal finally decided by the model.

[0123] The above description of the disclosed embodiments enables a person skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to these embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An electromagnetic spectrum radiation source modulation mode intelligent identification method based on unbalanced learning, characterized in that, The method comprises the following steps: a training phase and an inference phase of an intelligent identification model of a spatial electromagnetic spectrum radiation source modulation mode, wherein the intelligent identification model adopts a deep neural network model. The training phase of the model comprises the following steps: (1) establishing a wide-area spatial electromagnetic spectrum database comprising electromagnetic signal data to generate a data set required for model training; (2) converting the input modulation signal into an IQ signal form, and generating a data set required for training the model in the form of a double-channel sequence; (3) performing normalization preprocessing on the input data feature values, converting all data labels into one-hot encoding label form, and then dividing the data set into a training set, a validation set and a test set; (4) before each training round of the model, inputting the validation set data into the model, calculating a soft confusion matrix according to the model output and the validation set one-hot encoding label, and extracting the imbalance features of the data and the model; (5) updating the enhanced encoding generation matrix of the training set based on the soft confusion matrix, multiplying the enhanced encoding generation matrix and the training set one-hot encoding label to obtain the training set enhanced encoding label; (6) inputting the training set data into the model, calculating a loss function according to the model output and the training set enhanced encoding label obtained in the previous step, and then performing error back propagation based on the loss function to obtain the gradient of the loss function on each weight of the model, and updating the weight according to the gradient descent method; (7) repeating steps (4) to (6) in each training round until the training is completed, evaluating the model by using the test set data, and finally obtaining an electromagnetic spectrum radiation source modulation mode intelligent identification model; The inference phase of the model comprises the following steps: receiving electromagnetic signal data, performing IQ conversion and normalization preprocessing on the electromagnetic signal data, and then inputting the data into the model obtained after training, and automatically determining the modulation mode of the radiation source according to the model output; In step (4), the soft confusion matrix S is calculated as follows: S = L oT Y where L o is a matrix consisting of one-hot encoded labels of all examples in the set, Y is a matrix consisting of model output values based on the examples of the set, whose elements take values as follows: y i,q = P(q | x i ) The elements of the soft confusion matrix S are defined as follows: where P(q | x i ) denotes the probability that example i belongs to class q as predicted by the model, D p is the set of indices of all examples belonging to class p, i.e. i e D p denotes that example i belongs to class p, s p,q denotes the element in the p-th row and q-th column of the soft confusion matrix S; In step (5), the enhanced encoding generation matrix is represented as follows: G = (1 - μ)G * + μ(I - ∈C) where G * is the old value of the enhanced encoding generation matrix in a training round, which is initialized as an N-by-N identity matrix in the first training round; hyperparameters μ and ∈ are the update rate and the enhancement rate, respectively, I is an N-by-N identity matrix, N is the total number of classes in the dataset, and C is the cost matrix, which is obtained based on a soft confusion matrix of a validation set: C = S'-I Wherein S' is the row normalization of the soft confusion matrix, and the elements are: Training set augmented encoding labels The calculation is as follows: wherein, is a matrix composed of one-hot encoded labels of all examples in the training set, and the matrix multiplication is defined over the real number field.

2. The method of claim 1, wherein, In step (2), the modulation signal is converted into an IQ signal form, wherein I is the in-phase component and Q is the quadrature component.

3. The method of claim 2, wherein the method is characterized by, The mathematical relationship between the modulation signal and the IQ signal is as follows: x(t) = I(t) cos(2πf c t) + Q(t) sin(2πf c t) where x(t) is the modulating signal, I(t) and Q(t) are the I and Q signals, respectively, f c is the carrier frequency and t is time.

4. The method of claim 1, wherein the method is characterized by, In step (3), the input data feature values are normalized and preprocessed, and the value range is scaled to the interval [-1, 1]; the data labels are converted into one-hot encoding form, and if a sample belongs to the pth class, its one-hot encoding label is described as: wherein, l o is a one-hot encoded label, N is the total number of classes in the dataset, takes the following values: where p is the index of the class to which the example belongs, represents the l o th element of the qth column of the matrix 5. The method of claim 1, wherein, In step (6), the cross-entropy loss function calculated based on the enhanced encoding label is as follows: wherein, is an element of the training set augmented encoding label , N is the total number of classes in the dataset, y i,q is an element of the model output vector, q is the class number.

6. The method of claim 1, wherein the method is characterized by: In the model inference phase, the output of the model is a prediction vector, and each element of the vector represents the prediction vector of the input signal modulation mode belonging to the class: y = [y0, y1, y2,..., y q ,…,y N-1 ] y q = P(q | x) Wherein x represents the model input, y represents the model output, P(q|x) represents the model prediction probability that the input signal modulation mode belongs to the class q, and the modulation decision follows the criterion of the maximum prediction probability: Wherein, p is the modulation mode category of the input radiation source signal finally judged by the model.

Images