A method for generating delta modulation recognition model based on deep learning
Through the incremental modulation recognition model generation method based on deep learning, the recognition model is updated step by step, which solves the recognition problem caused by the variety of modulated signals in wireless communications, and realizes the high-accuracy recognition and real-time update capabilities in non-cooperation communications , adapt to practical application needs in complex environments.
Patent Information
- Application Number
- CN202011536524.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2020-12-23
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2040-12-23
AI Technical Summary
The prior art is difficult to quickly identify and update modulated signals in wireless communications, especially in complex environments with varying types of modulated signals, which cannot meet the high accuracy recognition requirements in non-cooperative communications.
The incremental modulation recognition model generation method based on deep learning is adopted, and the recognition model is updated step by step through data processing and incremental learning modules. The neural network and distillation technology are used to adapt to the changes in the types of modulated signals and achieve high accuracy recognition.
In non-cooperative communication, timely identification and high accuracy recognition of modulated signals are realized, which can adapt to the diversity of modulated signals and meet the real-time and performance requirements of practical applications.
Smart Images

Figure CN114662552B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a method for generating an incremental modulation recognition model based on deep learning, and belongs to the field of wireless communication and non-cooperative communication. Background Art
[0002] Wireless communication plays an increasingly important role in both the civilian and military fields. Therefore, the study of wireless communication is of great importance for the improvement of civilian science and technology and the enhancement of national defense strength. Wireless communication uses electromagnetic wave radiation in space to transmit information. This open channel environment makes countermeasure reconnaissance in non-cooperative communication possible. In communication reconnaissance, radio reconnaissance equipment will search, detect, and identify the signals transmitted by the enemy's wireless communication equipment. In order to effectively intercept the information in the signal, it is first necessary to identify the modulation mode of the signal. However, in recent years, with the rapid development of wireless communication technology and signal processing technology, the modulation mode of communication signals has become more and more complex, and the types of modulation signals have become more and more, which makes communication reconnaissance face severe challenges.
[0003] Based on the background of communication countermeasure reconnaissance in non-cooperative communication, this paper takes digital modulated signals as the research object, and studies the method of generating modulation recognition models when the types of signals are increasing in stages. Domestic and foreign studies have shown that in actual communication systems, modulation recognition algorithms should take into account both performance and real-time requirements. They should not only be able to effectively and quickly identify the commonly used communication signal modulation types, but also have the function of timely adjustment and update to cope with changes in signal modulation methods. In addition, for communication reconnaissance systems, since the intercepted communication signals generally lack prior knowledge, in order to accurately demodulate the modulated signals, it is necessary to use blind signal processing technology in modern digital signal processing technology, which forces us to find a modulation recognition model that does not require prior knowledge.
[0004] In recent years, many papers have studied the modulation recognition technology of digital modulated signals. Most of them design recognition models under the premise that the type of modulation signal is determined. The author Tian Xiaodi studied the problem of accurate recognition of communication signal modulation mode in the paper "Research on Communication Signal Modulation Pattern Recognition Method Based on Deep Learning [D]" [Changchun: Jilin University, 2020.], but the designed classification model can only be used when the type of modulation signal is fixed. It does not consider the impact of the increasing number of modulation signals and cannot meet the requirements of timely updates. Summary of the invention
[0005] In order to overcome the defects and shortcomings of the prior art, the present invention proposes a method for generating an incremental modulation recognition model based on deep learning, so as to timely identify the modulated signal with high accuracy in non-cooperative communication.
[0006] To achieve the above object, the technical solution of the present invention is as follows:
[0007] A method for generating an incremental modulation recognition model based on deep learning, the model includes a 0th stage and subsequent stages arranged in sequence, wherein the 0th stage includes a data processing module and an original network training module; each subsequent stage includes a data processing module and an incremental learning module; wherein the data processing module in each stage is connected to the original network training module or the incremental learning module, and the specific steps of the method are as follows:
[0008] 1) Data processing;
[0009] (1) The data processing module maps the collected modulated signal samples into a constellation diagram to obtain a sample set where x j represents the sample data in the i-th stage, S i represents the total number of samples in the i-th stage;
[0010] (2) Based on the sample set For each sample x j Type, add label y j , generate the i-th stage dataset Represents x j is an n-dimensional vector, Represents y j It is M i dimensional vector, where n represents the dimension of the sample, M i Indicates D i The number of types of samples in , let D i,m Yes D i The set of data of the mth category in ,
[0011] 2) Original network training:
[0012] ① The original network training module receives data from the data processing module and first initializes the memory data set P0 of stage 0. In stage 0, there are no samples in the memory data set, so it is set to an empty set. Right now
[0013] ② Initialize the recognition model according to the number of categories M0 contained in the 0-stage data set D0; use the neural network as the classification model, and set the model of the i-th stage to be Among them, L i is the number of layers of the neural network at stage i, θ i represents the parameter set of the neural network at stage i, Represents the 0th to the Lth layer of the neural network in stage i i The mapping relationship of the layer, that is, the mapping relationship from input to output, N lrepresents the output dimension of the lth layer, Represents a vector of dimension N0 to a vector of dimension Vector map of , symbols Represents the conversion of vector dimensions; represents the input of the neural network at stage i, where the symbol ∈ represents the belonging relationship in the set, and r i,0 After the neural network model φ i The output of the neural network in stage i can be obtained Therefore, the initial model at stage 0 is
[0014] ③ Input the samples in D0 into the initial model φ of stage 0 0,0 ,φ 0,0 The output is The loss at stage 0 is:
[0015]
[0016] ④ Use the SGD algorithm to train the model, minimize the loss L0 of stage 0, and generate a trained modulation recognition model φ0;
[0017] 3) Incremental learning;
[0018] a) The incremental learning module receives the data set D from the data processing module i , the input data set D of the previous stage i-1 and the recognition model φ output from the previous stage i-1 Identification model of the next generation cost stage i , according to the input φ i-1 and D i-1 , for the memory dataset P of stage i-1 i-1 Update and output the updated memory dataset P i , the specific steps of the update are as follows:
[0019] (1) From φ i-1 Separate the feature extraction part of the neural network from the L i-1 The first L layers of the neural network i-1 -2 layers are regarded as feature extractors, and the last fully connected layer of the neural network, i.e., the Lth i-1 -1 layer and softmax layer, namely the Lth i-1 The layer is used for classification, and the feature extractor only has φ i-1 Middle front L i -2 layer parameters Feature extractor at stage i
[0020] (2) Data set D i-1Each type of data input Get D i-1 The mean vector of each type of data in , where the mean vector of the mth type of data is: Finally, the mean vector set is obtained
[0021] (3) From D i-1 Each type of data in selects k data with the mean vector μ i-1,m The closest samples are combined to obtain a representative sample set Among them, x j,m Indicates D i-1 The jth sample of the mth class of data;
[0022] (4) Collect representative samples Join the memory dataset P i-1 Get the memory data set P i , let λ=M0+...+M i-1 represents the total data types arriving from stage 0 to stage i-1, P i The total number of samples contained in is: S P,i =λ×k;
[0023] b) P i The samples and D i Mix the samples in to get the mixed data set ε i ;
[0024] c) According to the input D i The number of types included M i , modify φ i-1 The number of output nodes in the last layer Let it be λ+M i , and obtain the initial model φ of stage i i,0 ;
[0025] d) Modify φ i,0 , in φ i,0 Insert a layer before the last layer to perform a linear transformation operation. The transformation coefficient depends on the temperature parameter T. The modified model is represents the neural network used in the i-th stage under the temperature parameter T, with L i +1 layer, the output of the last layer is used express;
[0026] e) Change ε i The samples in are sent to the model and output the value As a distillation label Join Epsilon i , generate a new mixed dataset It contains k samples of each type of data from stage 0 to stage i-1 and D i All samples in;
[0027] f) Sample x j Input Model The model can be obtained The last layer output and logits output Output logits Send it to the softmax function for normalization, and we can get For sample x j The predicted probability y pred,j , Therefore, ε′ i The set of all samples in Input Model In the neural network, we can get The last layer output and right The predicted probability set
[0028] g) Calculate the loss of stage i:
[0029] L i (θ i )=αL hard +βL soft (2)
[0030] in, is the cross entropy loss, is the distillation loss, α and β represent the weights of the cross entropy loss and distillation loss respectively, and L i (θ i ) is called cross-distillation loss;
[0031] h) Use the SGD algorithm to minimize formula (2) to train the model and obtain the neural network trained under temperature parameter T in stage i delete The penultimate layer in generates the modulation recognition model φ i .
[0032] The beneficial effects of the present invention are as follows: it can take into account the impact of an increasing number of types of modulation signals, meet the requirements of timely updating, and timely identify the modulation signals with high accuracy in non-cooperative communication, which is closer to practical applications. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] Figure 1 It is a schematic block diagram of the system of the present invention. DETAILED DESCRIPTION
[0034] The present invention will be further described below with reference to the accompanying drawings and embodiments, but is not limited thereto.
[0035] Example:
[0036] A method for generating delta modulation recognition models based on deep learning, such as Figure 1 As shown, the model includes stage 0 and subsequent stages arranged in sequence, wherein stage 0 includes a data processing module and an original network training module; each subsequent stage includes a data processing module and an incremental learning module; wherein the data processing module in each stage is connected to the original network training module or the incremental learning module, and the specific steps of the method are as follows:
[0037] 1) Data processing;
[0038] (1) The data processing module maps the collected modulated signal samples into a constellation diagram to obtain a sample set where x j represents the sample data in the i-th stage, S i represents the total number of samples in the i-th stage;
[0039] (2) Based on the sample set For each sample x j Type, add label y j , generate the i-th stage dataset Represents x j is an n-dimensional vector, Represents y j It is M i dimensional vector, where n represents the dimension of the sample, M i Indicates D i The number of types of samples in , let D i,m Yes D i The set of data of the mth category in ,
[0040] 2) Original network training:
[0041] ① The original network training module receives data from the data processing module and first initializes the memory data set P0 of stage 0. In stage 0, there are no samples in the memory data set, so it is set to an empty set. Right now
[0042] ② Initialize the recognition model according to the number of categories M0 contained in the 0-stage data set D0; use the neural network as the classification model, and set the model of the i-th stage to be Among them, L i is the number of layers of the neural network at stage i, θ i represents the parameter set of the neural network at stage i, Represents the 0th to the Lth layer of the neural network in stage i i The mapping relationship of the layer, that is, the mapping relationship from input to output, N l represents the output dimension of the lth layer, Represents a vector of dimension N0 to a vector of dimension Vector map of , symbols Represents the conversion of vector dimensions; represents the input of the neural network at stage i, where the symbol ∈ represents the belonging relationship in the set, and r i,0 After the neural network model φ i The output of the neural network in stage i can be obtained Therefore, the initial model at stage 0 is
[0043] ③ Input the samples in D0 into the initial model φ of stage 0 0,0 ,φ 0,0 The output is The loss at stage 0 is:
[0044]
[0045] ④ Use the SGD algorithm to train the model, minimize the loss L0 of stage 0, and generate a trained modulation recognition model φ0;
[0046] 3) Incremental learning;
[0047] a) The incremental learning module receives the data set D from the data processing module i , the input data set D of the previous stage i-1 and the recognition model φ output from the previous stage i-1 Identification model of the next generation cost stage i , according to the input φ i-1 and D i-1 , for the memory dataset P of stage i-1 i-1 Update and output the updated memory dataset P i , the specific steps of the update are as follows:
[0048] (1) From φ i-1 Separate the feature extraction part of the neural network from the L i-1 The first L layers of the neural network i-1 -2 layers are regarded as feature extractors, and the last fully connected layer of the neural network, i.e., the Lth i-1 -1 layer and softmax layer, namely the Lth i-1 The layer is used for classification, and the feature extractor only has φ i-1 Middle front L i -2 layer parameters Feature extractor at stage i
[0049] (2) Data set D i-1 Each type of data input Get D i-1 The mean vector of each type of data in , where the mean vector of the mth type of data is: Finally, the mean vector set is obtained
[0050] (3) From D i-1 Each type of data in selects k data with the mean vector μ i-1,m The closest samples are combined to obtain a representative sample set Among them, x j,m Indicates D i-1 The jth sample of the mth class of data;
[0051] (4) Collect representative samples Join the memory dataset P i-1 Get the memory data set P i , let λ=M0+...+M i-1 represents the total data types arriving from stage 0 to stage i-1, P i The total number of samples contained in is: S P,i =λ×k;
[0052] b) P i The samples and D i Mix the samples in to get the mixed data set ε i ;
[0053] c) According to the input D i The number of types included M i , modify φ i-1 The number of output nodes in the last layer Let it be λ+M i , and obtain the initial model φ of stage i i,0 ;
[0054] d) Modify φ i,0 , in φ i,0 Insert a layer before the last layer to perform a linear transformation operation. The transformation coefficient depends on the temperature parameter T. The modified model is represents the neural network used in the i-th stage under the temperature parameter T, with L i +1 layer, the output of the last layer is used express;
[0055] e) Change ε i The samples in are sent to the model and output the value As a distillation label Join Epsilon i , generate a new mixed dataset It contains k samples of each type of data from stage 0 to stage i-1 and D i All samples in;
[0056] f) Sample x j Input Model The model can be obtained The last layer output and logits output Output logits Send it to the softmax function for normalization, and we can get For sample x j The predicted probability y pred,j , Therefore, ε′ i The set of all samples in Input Model In the neural network, we can get The last layer output and right The predicted probability set
[0057] g) Calculate the loss of stage i:
[0058] L i (θ i )=αL hard +βL soft (2)
[0059] in, is the cross entropy loss, is the distillation loss, α and β represent the weights of the cross entropy loss and distillation loss respectively, and L i (θ i ) is called cross-distillation loss;
[0060] h) Use the SGD algorithm to minimize formula (2) to train the model and obtain the neural network trained under temperature parameter T in stage i delete The penultimate layer in generates the modulation recognition model φ i .
Claims
1. A method for generating an incremental modulation recognition model based on deep learning, the model includes a 0th stage and subsequent stages arranged in sequence, wherein the 0th stage includes a data processing module and an original network training module; each subsequent stage includes a data processing module and an incremental learning module; wherein the data processing module in each stage is connected to the original network training module or the incremental learning module, and the specific steps of the method are as follows: 1) Data processing; (1) The data processing module maps the collected modulated signal samples into a constellation diagram to obtain a sample set where x j represents the sample data in the i-th stage, S i represents the total number of samples in the i-th stage; (2) Based on the sample set For each sample x j Type, add label y j , generate the i-th stage data set Represents x j is an n-dimensional vector, Represents y j It is M i dimensional vector, where n represents the dimension of the sample, M i Indicates D i The number of types of samples in , let D i,m Yes D i The set of m-th type of data, 2) Original network training: ① The original network training module receives data from the data processing module and first initializes the memory data set P0 of stage 0. In stage 0, there are no samples in the memory data set, so it is set to an empty set. Right now ② Initialize the recognition model according to the number of categories M0 contained in the 0-stage data set D0; use the neural network as the classification model, and set the model of the i-th stage to be in, L i is the number of layers of the neural network at stage i, θ i represents the parameter set of the neural network at stage i, Represents the 0th to the Lth layer of the neural network in stage i i The mapping relationship of the layer, that is, the mapping relationship from input to output, N l represents the output dimension of the lth layer, Represents a vector of dimension N0 to a vector of dimension Vector map of , symbols Represents the conversion of vector dimensions; represents the input of the neural network at stage i, where the symbol ∈ represents the belonging relationship in the set, and r i,0 After the neural network model φ i The output of the neural network in stage i can be obtained Therefore, the initial model at stage 0 is ③ Input the samples in D0 into the initial model φ of stage 0 0,0 ,φ 0,0 The output is The loss at stage 0 is: ④ Use the SGD algorithm to train the model, minimize the loss L0 of stage 0, and generate a trained modulation recognition model φ0; 3) Incremental learning; a) The incremental learning module receives the data set D from the data processing module i , the input data set D of the previous stage i-1 and the recognition model φ output from the previous stage i-1 Identification model of the next generation cost stage i , according to the input φ i-1 and D i-1 , for the memory dataset P of stage i-1 i-1 Update and output the updated memory dataset P i , the specific steps of the update are as follows: (1) From φ i-1 Separate the feature extraction part of the neural network from the L i-1 The first L layers of the neural network i-1 -2 layers are regarded as feature extractors, and the last fully connected layer of the neural network, i.e., the Lth i-1 -1 layer and softmax layer, namely the Lth i-1 The layer is used for classification, and the feature extractor only has φ i-1 Middle front L i -2 layer parameters Feature extractor at stage i (2) Data set D i-1 Each type of data input Get D i-1 The mean vector of each type of data in , where the mean vector of the mth type of data is: Finally, the mean vector set is obtained (3) From D i-1 Each type of data in selects k data with the mean vector μ i-1,m The closest samples are combined to obtain a representative sample set Among them, x j,m Indicates D i-1 The jth sample of the mth class of data; (4) Collect representative samples Join the memory dataset P i-1 Get the memory data set P i , let λ=M0+...+M i-1 represents the total data types arriving from stage 0 to stage i-1, P i The total number of samples contained in is: S P,i =λ×k; b) P i The samples and D i Mix the samples in to get the mixed data set ε i ; c) According to the input D i The number of types included M i , modify φ i-1 The number of output nodes in the last layer Let it be λ+M i , and obtain the initial model φ of stage i i,0 ; d) Modify φ i,0 , in φ i,0 Insert a layer before the last layer to perform a linear transformation operation. The transformation coefficient depends on the temperature parameter T. The modified model is represents the neural network used in the i-th stage under the temperature parameter T, with L i +1 layer, the output of the last layer is used express; e) Change ε i The samples in are sent to the model and output the value As a distillation label Join ε i , generate a new mixed dataset It contains k samples of each type of data from stage 0 to stage i-1 and D i All samples in; f) Sample x j Input Model The model can be obtained The last layer output and logits output Output logits Send it to the softmax function for normalization, and we can get For sample x j The predicted probability y pred,j , Therefore, ε′ i The set of all samples in Input Model In the neural network, we can get The last layer output and right The predicted probability set g) Calculate the loss of stage i: L i (i i )=αL hard +βL soft (2) in, is the cross entropy loss, is the distillation loss, α and β represent the weights of the cross entropy loss and distillation loss respectively, and L i (θ i ) is called cross-distillation loss; h) Use the SGD algorithm to minimize formula (2) to train the model and obtain the neural network trained under temperature parameter T in stage i delete The penultimate layer in generates the modulation recognition model φ i .
Citation Information
Patent Citations
Method for identifying subcarrier modulation constellation diagram of wireless optical communication under atmospheric turbulence
CN108052956A
Modulation signal incremental learning method based on fixed features
CN111210000A
Training of deep neural networks on the basis of distributions of paired similarity measures
WO2018124918A1
