An intelligent spectrum sensing method with environmental adaptability

Through the adversarial learning method of generating adversarial networks, the spectrum state characteristics that are decoupled from the environment are learned, and the problem of insufficient generalization ability of intelligent spectrum perception in complex noise environments is solved, efficient spectrum perception in unknown environments is realized, and application is improved The breadth of

CN114118151BActive Publication Date: 2024-11-01THE 54TH RESEARCH INSTITUTE OF CHINA ELECTRONICS TECHNOLOGY GROUP CORPORATION

Patent Information

Application Number
CN202111410632.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-11-25
Publication Date
2024-11-01
Estimated Expiration
2041-11-25

AI Technical Summary

Technical Problem

The existing intelligent spectrum perception method is difficult to generalize to unknown environments in complex noise environments, and requires a large amount of labeled data for retraining, which limits its wide range of applications.

Method used

Adversarial learning method that generates adversarial networks, and learns spectrum state features decoupled from the training data through neural networks, including feature extractors, spectral state classifiers and domain discriminators. Adversarial learning makes the output of feature extractor not include Domain-related information to achieve environmental adaptability of spectrum perception.

Benefits of technology

In the context of lack of prior information, online adaptive and agile spectrum perception is achieved through offline one-time learning, which improves the generalization ability of intelligent spectrum perception and can work effectively in unknown environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114118151B_ABST
    Figure CN114118151B_ABST
Patent Text Reader

Abstract

The present invention relates to a wireless communication technology, and in particular to a spectrum sensing method based on data-driven. The specific process is: in the training stage, a neural network is used to learn the spectrum state characteristics that are decoupled from the environment from the training data sample set; in the test stage, the test data is input into the trained neural network to obtain the spectrum sensing classification result. The neural network uses the idea of ​​generating adversarial networks to extract the characteristics of decoupling from the environment through adversarial learning. In the absence of prior information, the present invention can use a limited set of training samples in a known domain to learn spectrum state characteristics that are independent of the environment through offline one-time learning, thereby completing online adaptive and agile spectrum sensing, realizing spectrum sensing capabilities with environmental adaptability, and effectively improving the generalization ability of intelligent spectrum sensing methods.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a wireless communication technology, and in particular to a spectrum sensing method based on data driving. Background Art

[0002] The rapid development of wireless communication technology has led to an increasingly prominent contradiction between the demand for high-speed and high-quality wireless data services and the saturated spectrum resources. However, the spectrum resources of many existing wireless systems are idle to varying degrees in time and space. Spectrum sensing can enable wireless communication devices to have efficient electromagnetic environment perception capabilities, dynamically coordinate various wireless communication devices, and provide them with access to unused idle spectrum resources, thereby increasing the access volume of devices and services and improving spectrum utilization. It provides a feasible solution. The two key indicators of spectrum sensing are timeliness and accuracy. In order to ensure the normal operation of dynamic spectrum access, it is necessary to ensure that the perception results of the entire communication system can be obtained quickly and accurately. Therefore, the quality of spectrum sensing performance is the premise and necessary condition for the efficient use of spectrum resources. Traditional spectrum sensing methods include energy detection algorithm, matched filter detection algorithm, cyclostationary feature detection algorithm, etc. These traditional algorithms are mainly based on thresholds for judgment. The inaccuracy of threshold calculation is the main reason for the decline in detection performance.

[0003] With the deep integration of the Internet of Things, big data, cloud computing and intelligent algorithms, artificial intelligence, as the most disruptive and transformative basic technology, has shown a very broad development space and is empowering all aspects of scientific and technological innovation, economic production and social life. In classification tasks, machine learning methods in artificial intelligence can not only achieve nonlinear classification boundaries, but also easily integrate features with different distributions or attributes for classification and discrimination, which provides an innovative solution for spectrum perception. When the relationship between operating parameters and electromagnetic environment is unknown, machine learning methods with autonomous learning capabilities can discover the mapping relationship between data and spectrum occupancy status from data, solving spectrum perception problems in the absence of prior information. It constructs a statistical model through a data set, and after continuous iterative optimization of the objective function, it finally trains a model that matches the actual system, and then performs spectrum status prediction and analysis based on this model, which can effectively improve the perception performance of a specific environment.

[0004] However, the working mode of offline learning and online work of machine learning methods has a limitation. The domain characteristics of data reflect the distribution characteristics of data. Different domains of data have different distributions of data. Machine learning methods require that the input data for online testing and offline learning must obey the same distribution (domain), or the distribution of online test data belongs to one of the distributions of offline learning training data, otherwise the model performance will decline. For the application of machine learning methods in spectrum sensing, the distribution of data is different in different communication environments. This requires that the communication environment where the data is located during online testing is one of the communication environments collected by the offline training data set. Otherwise, the perception ability of the model will be greatly reduced. However, the communication environment of actual applications is often complex and changeable. If the model is retrained in actual applications, it is difficult for the neural network to train the model with a large amount of labeled data. Therefore, this problem seriously affects the widespread application of machine learning methods in spectrum sensing. How to effectively generalize intelligent spectrum sensing to unknown environments in complex noise environments is a key element of intelligent spectrum sensing. Summary of the invention

[0005] Aiming at the environmental adaptability requirements of intelligent spectrum sensing, the present invention utilizes the adversarial learning capability of generative adversarial networks to provide an intelligent spectrum sensing technology with environmental adaptability, thereby promoting the application of the intelligent spectrum sensing technology.

[0006] The technical solution adopted by the present invention is:

[0007] An intelligent spectrum sensing method with environmental adaptability includes the following processes:

[0008] In the training phase, the neural network is used to learn the spectrum state characteristics of the environment decoupling from the training data sample set;

[0009] In the testing phase, the test data is input into the trained neural network to obtain the spectrum sensing classification results.

[0010] Furthermore, the spectrum state feature that is decoupled from the environment refers to a feature of the spectrum state that does not change with changes in the environment.

[0011] Furthermore, the neural network utilizes the idea of ​​generative adversarial networks to extract features that are decoupled from the environment through adversarial learning.

[0012] Furthermore, the neural network at least includes a feature extractor E, a spectrum state classifier C and a domain identifier D.

[0013] Furthermore, the feature extractor E is used to extract data features carrying spectrum state information from the input data, the spectrum state classifier C is used to classify the spectrum state of the data features and determine whether the spectrum state is occupied or idle, and the domain identifier D is used to identify the domain category of the data features and determine the domain from which the data features come.

[0014] Furthermore, the input data is the original data of the received signal or the data after preprocessing.

[0015] Furthermore, the feature extractor E is composed of multiple convolutional layers; the spectral state classifier C is composed of multiple fully connected layers; and the domain discriminator D is composed of multiple fully connected layers.

[0016] Furthermore, the spectrum state category is represented by one-hot encoding; the domain category is represented by one-hot encoding.

[0017] Furthermore, the feature extractor and the domain discriminator are in an adversarial learning relationship; the ultimate goal of adversarial learning is to make the features extracted by the feature extractor contain no domain-related information, so that the domain discriminator cannot determine the domain category of the features extracted by the feature extractor, thereby achieving decoupling of the features extracted by the feature extractor from the environment.

[0018] Furthermore, when the neural network is trained, the parameters θ of the feature extractor E are updated by jointly minimizing the loss function of EC and maximizing the loss function of ED. e , update the parameters θ of the spectrum state classifier C by minimizing the loss function of EC c , update the parameters θ of the domain discriminator D by maximizing the loss function of ED d .

[0019] Furthermore, the loss function is obtained based on the cross entropy.

[0020] Furthermore, jointly minimizing the loss function of EC and maximizing the loss function of ED means introducing a gradient reversal layer between the feature extractor E and the domain discriminator D, and jointly optimizing and updating the parameters of the feature extractor by controlling the hyperparameters of the gradient reversal layer.

[0021] Furthermore, the training data sample set consists of input data, spectrum state labels and environment labels, which can be expressed as: {Ψ, Z, T} = {(S 1 ,z 1 ,t 1 ),...,(S K ,z K ,t K )}, where Ψ=(S 1 , ..., S K ) represents K input data sets, Z=(z 1 ,...,zK ), z k ∈{0,1} represents the spectrum state label set, T=(t 1 ,...,t K ) represents a set of environment labels;

[0022] Furthermore, the loss function of EC is minimized and the loss function of ED is maximized as the loss function of the entire neural network model. The expression is as follows:

[0023] J(θ e ,θ c ,θ d )=J(θ e ,θ c )-λJ(θ e ,θ d )

[0024] Where λ is the introduced hyperparameter, λ>0; J(θ e ,θ c ) is the loss function of the feature extractor E-spectral state classifier C, J(θ e ,θ d ) is the loss function of feature extractor E-domain discriminator D.

[0025] Furthermore, the parameters of the feature extractor E, the spectrum state classifier C and the domain discriminator D are updated by constructing an optimization objective function;

[0026] The constructed optimization objective function is as follows:

[0027]

[0028] In the formula, and They correspond to the optimization parameters of feature extractor E, spectrum state classifier C and domain discriminator D respectively;

[0029] By continuously iteratively optimizing the objective function, the optimal parameters of the feature extractor E, spectral state classifier C and domain discriminator D are obtained.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] In the absence of prior information, the present invention can utilize a limited set of training samples in the known domain to learn spectrum state characteristics that are independent of the environment through offline one-time learning, thereby completing online adaptive and agile spectrum sensing, realizing spectrum sensing capabilities with environmental adaptability, and effectively improving the generalization ability of intelligent spectrum sensing methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1This is a spectrum sensing neural network model diagram based on adversarial learning in an embodiment of the present invention. DETAILED DESCRIPTION

[0033] The present invention will be further explained below in conjunction with the accompanying drawings and specific embodiments.

[0034] Embodiment: Considering the model of single-user multi-antenna, different environments (domains) are represented by considering different signal-to-noise ratio (SNR) scenarios. In the embodiment, the neural network input data is preprocessed data, and the covariance matrix of the received signal is used here. For communication data under multiple SNR scenarios, the spectrum sensing neural network based on adversarial learning proposed by the method of the present invention is used to realize intelligent spectrum sensing that can adapt to unknown SNR scenarios. The specific implementation method is as follows:

[0035] 1. Training Dataset

[0036] The received signals under multiple SNRs and the labels of their corresponding spectrum states are collected, and the sampling covariance matrix of the received signals is calculated as training samples.

[0037] Assuming that the number of SNR scenes of the collected training data is V, the parameter t represents the SNR label corresponding to the training sample, and z represents the spectrum state label corresponding to the training sample, then the entire training data set can be expressed as:

[0038]

[0039] Among them, K represents the number of training samples, represents the K sampling covariance matrix, Z = (z 1 ,...,z K ), z k ∈{0,1} represents the spectrum state label set, T=(t 1 ,...,t K ) represents the SNR label set, t k It can be indicated by the subscript of the SNR value.

[0040] 2. Offline training

[0041] The training data set is input into the neural network proposed by the method of the present invention, and the neural network is trained offline. Figure 1 As shown, it includes feature extractor E, spectrum state classifier C and domain discriminator D. The model parameters of E, C and D are represented by θ e ,θ c and θ d express.

[0042] The feature extractor E extracts data features f from the training data sample set, and the data features f carry spectrum state information; the spectrum state classifier C classifies the spectrum state of the output f of the feature extractor E and determines whether the spectrum state is in an occupied state (H 1 ) or idle state (H 0 ); The domain identifier D performs domain category identification on the output f of the feature extractor E to determine which domain f comes from.

[0043] The feature extractor E and the domain discriminator D are in an adversarial learning relationship. The ultimate goal of adversarial learning is to make the output f of the feature extractor E contain no domain-related information, so that D cannot determine the domain category of f, and to decouple the feature f extracted by the feature extractor E from the environment, thereby removing the correlation between the feature f and the environment.

[0044] In the offline training stage, on the one hand, the loss function of EC is minimized to ensure that the feature f extracted by E can correctly predict the spectrum state; on the other hand, the adversarial learning between E and D is used to reduce the dependence of feature f on SNR. Specifically, the parameters θ of the feature extractor E are updated by jointly minimizing the loss function of EC and maximizing the loss function of ED. e , by minimizing the loss function of EC to update the parameters θ of the spectrum state classifier C c , update the parameters θ of the domain discriminator D by minimizing the loss function of ED d The optimization objective function is as follows:

[0045]

[0046]

[0047] In the formula, and They correspond to the optimization parameters of the feature extractor E, the spectrum state classifier C and the domain discriminator D respectively; by continuously iteratively optimizing the objective function, the optimal parameters of the feature extractor E, the spectrum state classifier C and the domain discriminator D are obtained.

[0048] J(θ e ,θ c ,θ d ) represents the loss function of the entire model, which is expressed as follows:

[0049] J(θ e ,θ c ,θ d )=J(θ e ,θ c )-λJ(θ e ,θ d ),

[0050] In the formula, J(θ e ,θ c ) represents the loss function of the network EC, J(θ e ,θ d ) represents the loss function of the network ED. The parameter λ>0 is a hyperparameter that introduces the gradient reverse layer between E and D to balance the two optimization objectives. During forward propagation, the gradient reverse layer does not perform any operation and directly passes the input to the next layer. During backward propagation, the gradient reverse layer multiplies the propagated gradient by the factor -λ and then passes it to the next layer.

[0051] The loss function is defined by cross entropy, assuming t k is encoded as a one-hot vector δ k =[δ 1 ,δ 2 ,...,δ V ], the output of network D is represented by vector To express, the loss functions of network EC and network ED are as follows:

[0052]

[0053]

[0054] In the formula, θ e ,θ c and θ d They correspond to the parameters of the feature extractor E, the spectrum state classifier C and the domain discriminator D respectively; represents the kth sampling covariance matrix, Indicates that the spectrum is in an occupied state H 1 In this case, input data The predicted output of the spectrum state after network EC, Indicates that the spectrum is idle 0 In this case, input data The predicted output of the spectrum state after network EC, z k represents the spectral state label corresponding to the kth training sample (here: covariance matrix), represents the SNR label of the kth training sample in the i-th SNR scenario, represents the prediction output of the domain discriminator D for the kth training sample in the i-th SNR scenario.

[0055] The hyperparameter value of the gradient reversal layer is an adaptive factor that changes gradually during the training process, as follows:

[0056]

[0057] In the formula, α is a variable that gradually changes from 0 to 1 as the training progresses.

[0058] 3. Online perception

[0059] During online detection, the trained network EC is used to predict the spectrum state. During detection, the sampling covariance matrix of the test data is first calculated, and then the trained model is input to obtain the classification result of the spectrum state classifier C to complete online real-time spectrum perception.

Claims

1. An intelligent spectrum sensing method with environmental adaptability, It is characterized in that The process includes: In the training phase, the neural network is used to learn the spectrum state characteristics of the environment decoupling from the training data sample set; In the testing phase, the test data is input into the trained neural network to obtain the spectrum sensing classification results; The neural network at least includes a feature extractor, a spectrum state classifier and a domain identifier; the feature extractor is used to extract data features carrying spectrum state information from input data, the spectrum state classifier is used to classify the spectrum state of the data features and determine whether the spectrum state is occupied or idle, and the domain identifier is used to identify the domain category of the data features and determine the domain from which the data features come; The spectrum state categories are represented by one-hot encoding; The domain category is represented by one-hot encoding. The feature extractor and the domain discriminator are in an adversarial learning relationship. The ultimate goal of adversarial learning is to make the features extracted by the feature extractor contain no domain-related information. Among them, when the neural network is trained, the parameters of the feature extractor are updated by jointly minimizing the loss function of the feature extractor-spectral state classifier and maximizing the loss function of the feature extractor-domain discriminator, the parameters of the spectrum state classifier are updated by minimizing the loss function of the feature extractor-spectral state classifier, and the parameters of the domain discriminator are updated by maximizing the loss function of the feature extractor-domain discriminator; The joint minimization of the loss function of the feature extractor-spectral state classifier and the maximization of the loss function of the feature extractor-domain discriminator refers to introducing a gradient reversal layer between the feature extractor and the domain discriminator, and jointly optimizing and updating the parameters of the feature extractor based on controlling the hyperparameters of the gradient reversal layer; The loss function of the joint minimization feature extractor-spectral state classifier and the maximization feature extractor-domain discriminator are used as the loss function of the entire neural network model, and the expression is as follows: J(θ e ,i c ,i d )=J(θ e ,i c )-λJ(θ e ,i d ) Where λ is the introduced hyperparameter, λ>0; θ e ,θ c and θ d They correspond to the parameters of the feature extractor, spectral state classifier, and domain discriminator respectively; J( θ e ,θ c ) is the loss function of the feature extractor-spectral state classifier, J(θ e ,θ d ) is the loss function of the feature extractor-domain discriminator; Among them, the loss function is defined by the cross entropy, assuming that t k is encoded as a one-hot vector δ k =[δ 1 ,δ 2 ,...,δ V ], the output of the domain discriminator is represented by the vector To express it, the loss functions of the feature extractor-spectral state classifier and the feature extractor-domain discriminator are as follows: In the formula, θ e ,θ c and θ d They correspond to the parameters of the feature extractor, spectrum state classifier and domain discriminator respectively, K represents the number of training samples, represents the kth sampling covariance matrix, Indicates that the spectrum is in an occupied state H 1 In this case, input data The predicted output of the spectrum state after the feature extractor-spectrum state classifier, Indicates that the spectrum is idle 0 In this case, input data The predicted output of the spectrum state after the feature extractor-spectrum state classifier, z k represents the spectral state label corresponding to the kth training sample, i.e. the covariance matrix, represents the SNR label of the kth training sample in the i-th SNR scenario, represents the prediction output of the domain discriminator for the kth training sample in the i-th SNR scenario; Among them, the hyperparameter value of the gradient reverse layer is an adaptive factor that gradually changes with the training process, as follows: In the formula, α is a variable that gradually changes from 0 to 1 as the training progresses; By constructing an optimization objective function, the parameters of the feature extractor, spectral state classifier, and domain discriminator are updated; the constructed optimization objective function is as follows: In the formula, and They correspond to the optimization parameters of the feature extractor, spectrum state classifier and domain discriminator respectively; By continuously iteratively optimizing the objective function, the optimal parameters of the feature extractor, spectral state classifier and domain discriminator are obtained.

2. The intelligent spectrum sensing method with environmental adaptability according to claim 1, It is characterized in that The spectrum state feature that is decoupled from the environment refers to the feature of the spectrum state that does not change with changes in the environment.

3. The intelligent spectrum sensing method with environmental adaptability according to claim 1, It is characterized in that The neural network utilizes the idea of ​​generative adversarial networks to extract features that are decoupled from the environment through adversarial learning.

4. The intelligent spectrum sensing method with environmental adaptability according to claim 1, It is characterized in that The input data is the original data of the received signal or the data after preprocessing.

5. The intelligent spectrum sensing method with environmental adaptability according to claim 1, It is characterized in that The feature extractor consists of multiple convolutional layers; the spectral state classifier consists of multiple fully connected layers; and the domain discriminator consists of multiple fully connected layers.

6. The intelligent spectrum sensing method with environmental adaptability according to claim 1, It is characterized in that The loss function is obtained based on cross entropy.

7. The intelligent spectrum sensing method with environmental adaptability according to claim 1, It is characterized in that The training data sample set consists of input data, spectrum state labels and environment labels, which can be expressed as: {Ψ, Z, Τ} = {(S 1 ,z 1 ,t 1 ),...,(S K ,z K ,t K )}, where Ψ=(S 1 , ..., S K ) represents K input data sets, Z=(z 1 ,...,z K ), z k ∈{0,1} represents the spectrum state label set, T=(t 1 ,...,t K ) represents a collection of environment tags.

Citation Information

Patent Citations

  • Interference avoidance communication model based on meta-learning and training method thereof

    CN110995382A

  • Information processing method, system and device

    CN111342984A

  • Semi-supervised modulation type identification method and device based on network structure feature induction and medium

    CN113420817A

  • Spectrum sensing method and system based on CGAN data enhancement

    CN113435263A

  • Limited data spectrum sensing method based on semi-supervised deep neural network

    CN113570032A

Cited By

  • Cellular spectrum illegal occupation identification method, system and device based on spectrum contour form, and medium

    CN122069525A