WO2020228141A1 - 基于内隐知识构建图卷积网络的电磁信号识别方法及装置 - Google Patents

基于内隐知识构建图卷积网络的电磁信号识别方法及装置 Download PDF

Info

Publication number
WO2020228141A1
WO2020228141A1 PCT/CN2019/098100 CN2019098100W WO2020228141A1 WO 2020228141 A1 WO2020228141 A1 WO 2020228141A1 CN 2019098100 W CN2019098100 W CN 2019098100W WO 2020228141 A1 WO2020228141 A1 WO 2020228141A1
Authority
WO
WIPO (PCT)
Prior art keywords
electromagnetic signals
implicit knowledge
graph
neural network
graph convolutional
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2019/098100
Other languages
English (en)
French (fr)
Inventor
杨昉
邹琮
潘长勇
王军
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Tsinghua University
Original Assignee
Tsinghua University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Tsinghua University filed Critical Tsinghua University
Publication of WO2020228141A1 publication Critical patent/WO2020228141A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/08—Feature extraction
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/12—Classification; Matching

Definitions

  • the feedback of this application belongs to the technical field of electromagnetic signal intelligent perception, and particularly relates to an electromagnetic signal recognition method and device based on implicit knowledge to construct a graph convolution network.
  • the feedback of this application aims to solve one of the technical problems in related technologies at least to a certain extent.
  • one purpose of the feedback of this application is to propose an electromagnetic signal recognition method based on implicit knowledge to construct a graph convolutional network.
  • This method can use the characteristics of each node to mine the relationship between nodes, and for each node The characteristics of electromagnetic signal categories are extracted deeply, and more information is used, and the generalization ability is stronger.
  • Another purpose of the feedback of this application is to propose an electromagnetic signal recognition device that constructs a graph convolutional network based on implicit knowledge.
  • the present application feedbacks that one embodiment of the present application proposes an electromagnetic signal recognition method based on implicit knowledge to construct a graph convolutional network, including:
  • S1 Acquire multiple electromagnetic signals, extract feature descriptions of the multiple electromagnetic signals, and generate a feature matrix as the first type of input information according to the feature descriptions of the multiple electromagnetic signals;
  • the electromagnetic signal recognition method of constructing graph convolutional network based on implicit knowledge of the embodiment of the present application feedback embodiment, constructing graph convolutional neural network based on implicit knowledge of electromagnetic signal can mine the relationship between nodes, and for each electromagnetic signal The feature of the category is extracted in depth, using more information, and the generalization ability is strong. At the same time, it makes full use of the prior knowledge of the signal, which increases the accuracy of the description of the relationship between the signals, improves the classification accuracy, and improves the performance of the neural network. Controllability and interpretability.
  • the electromagnetic signal recognition method for constructing graph convolutional network based on implicit knowledge of the foregoing embodiment feedback according to the present application may also have the following additional technical features:
  • the implicit knowledge includes: modulation type, carrier frequency, symbol rate, radio frequency characteristics, frequency spectrum characteristics, and channel characteristics.
  • the characteristics of the multiple electromagnetic signals are described as time-frequency characteristics
  • the feature descriptions of the multiple electromagnetic signals are extracted through short-time Fourier transform, wavelet transform and Hilbert-Huang transform.
  • the S1 further includes:
  • the dimensionality reduction processing of the time-frequency features is performed through a component analysis method, a linear discriminant analysis method and a local linear embedding method.
  • the S2 specifically includes:
  • the S3 further includes: evaluating the accuracy of the categories of the multiple electromagnetic signals output by the graph convolutional neural network by calculating cross entropy.
  • the present application feeds back another embodiment and proposes an electromagnetic signal recognition device based on implicit knowledge to construct a graph convolutional network, including:
  • the time-frequency feature extraction module is used to obtain multiple electromagnetic signals, extract feature descriptions of the multiple electromagnetic signals, and generate a feature matrix as the first type of input information according to the feature descriptions of the multiple electromagnetic signals;
  • the implicit knowledge graph construction module is used to mine the implicit knowledge of the multiple electromagnetic signals, and construct a graph structure based on the implicit knowledge of the electromagnetic signals according to the implicit knowledge as the second type of input information;
  • the electromagnetic signal classification module is used to build a graph convolutional neural network according to the first type of input information and the second type of input information, and use a weakly supervised learning method to perform cyclic and iterative training on the graph convolutional neural network, so that The graph convolutional neural network outputs the types of the multiple electromagnetic signals.
  • the electromagnetic signal recognition device for constructing graph convolutional network based on implicit knowledge of the embodiment of the present application feedback embodiment can mine the relationship between nodes, for each electromagnetic signal
  • the feature of the category is extracted in depth, using more information, and the generalization ability is strong.
  • it makes full use of the prior knowledge of the signal which increases the accuracy of the description of the relationship between the signals, improves the classification accuracy, and improves the performance of the neural network. Controllability and interpretability.
  • the electromagnetic signal recognition device for constructing graph convolutional network based on implicit knowledge of the foregoing embodiment feedback according to the present application may also have the following additional technical features:
  • the characteristics of the multiple electromagnetic signals are described as time-frequency characteristics
  • the implicit knowledge includes: modulation type, carrier frequency, symbol rate, radio frequency characteristics, frequency spectrum characteristics and channel characteristics.
  • time-frequency feature extraction module is also used for:
  • the dimensionality reduction processing of the time-frequency features is performed through a component analysis method, a linear discriminant analysis method and a local linear embedding method.
  • the electromagnetic signal classification module is also used to evaluate the accuracy of the categories of the multiple electromagnetic signals output by the graph convolutional neural network by calculating cross entropy.
  • Fig. 1 is a flow chart of an electromagnetic signal recognition method based on implicit knowledge to construct a graph convolutional network according to an embodiment of the present application
  • Fig. 2 is a schematic diagram of time-frequency images of different electromagnetic signals fed back according to an embodiment of the present application
  • Fig. 3 is a schematic diagram of electromagnetic signal propagation according to an embodiment of feedback of the present application.
  • FIG. 4 is a schematic diagram of a graph structure based on the implicit knowledge of electromagnetic signals according to an embodiment of the present application
  • FIG. 5 is a structural diagram of a graph convolutional neural network based on implicit knowledge of electromagnetic signals according to an embodiment of the present application
  • Fig. 6 is a flow chart of an electromagnetic signal recognition method for constructing a graph convolutional network based on implicit knowledge according to another embodiment of the present application;
  • Fig. 7 is a schematic structural diagram of an electromagnetic signal recognition device for constructing a graph convolutional network based on implicit knowledge according to an embodiment of the present application.
  • FIG. 1 is a flowchart of an electromagnetic signal recognition method based on implicit knowledge to construct a graph convolutional network according to an embodiment of the present application.
  • the electromagnetic signal recognition method of constructing graph convolutional network based on implicit knowledge includes the following steps:
  • Step S1 Acquire multiple electromagnetic signals, extract feature descriptions of the multiple electromagnetic signals, and generate a feature matrix as the first type of input information according to the feature descriptions of the multiple electromagnetic signals.
  • the feature descriptions of multiple electromagnetic signals are extracted in various ways to form a feature matrix as the first type of input information of the constructed graph convolutional neural network.
  • the feature descriptions of multiple electromagnetic signals extracted are time-frequency features, and there are many ways to extract feature descriptions, including but not limited to short-time Fourier Transform (Short Time Fourier Transform, STFT), Wavelet Transform (Wavelet Transform, WT) and Hilbert-Huang Transform (Hilbert-Huang Transform, HHT).
  • STFT Short Time Fourier Transform
  • WT Wavelet Transform
  • HHT Hilbert-Huang Transform
  • S1 also includes:
  • PCA Principal Component Analysis
  • LDA Linear Discriminant Analysis
  • LLE Locally Linear Embedding
  • short-time Fourier Transform is used to characterize electromagnetic signals as two-dimensional time-frequency distributions, which can be used as the feature description of nodes.
  • STFT short-time Fourier Transform
  • Figure 2 the STFT time-frequency images of four different electromagnetic signals are given, and it can be seen that the time-frequency distributions of different types of electromagnetic signals are obviously different.
  • the resolution accuracy the dimensionality of the time-frequency image is usually high, which will lead to a huge amount of calculation.
  • Step S2 mining the implicit knowledge of multiple electromagnetic signals, and constructing a graph structure based on the implicit knowledge of electromagnetic signals according to the implicit knowledge as the second type of input information.
  • implicit knowledge includes, but is not limited to, modulation type, carrier frequency, symbol rate, radio frequency characteristics, spectrum characteristics, and channel characteristics.
  • S2 specifically includes: mining the implicit knowledge of multiple electromagnetic signals to compare the correlation between different electromagnetic signals, generating an adjacency matrix based on the correlation between different electromagnetic signals, and constructing a graph structure based on the adjacency matrix as the second Class input information, where the graph nodes of the graph structure represent different electromagnetic signals.
  • the information contained in its nodes is the characteristic description of different electromagnetic signals, and the connection relationship between the nodes can be given by the adjacency matrix.
  • Figure 3 (a), Figure 3 (b), Figure 3 (c), Figure 3 (d), Figure 3 (e), Figure 3 (f) represent signal, modulation, digital-to-analog conversion , Channel, radio frequency, and antenna are further explanations of Figure 3(g).
  • each vertex is known class electromagnetic signals i ⁇ X 1 or unknown class described electromagnetic signal i ⁇ X 2 x i
  • the value of the adjacency matrix is between 0 and 1, which is determined by the implicit knowledge of the electromagnetic signal.
  • Step S3 build a graph convolutional neural network based on the first type of input information and the second type of input information, and use the weakly supervised learning method to perform cyclic and iterative training on the graph convolutional neural network, so that the graph convolutional neural network outputs multiple electromagnetic signals Category.
  • S3 also includes: evaluating the degree of consistency between the categories of the multiple electromagnetic signals output by the graph convolutional neural network and their actual categories by calculating cross entropy.
  • the condition for training suspension is: the absolute error of the cross entropy of two consecutive cyclic iterative training is less than the preset minimum Value, and the cross entropy of a single cycle training is less than the preset minimum value, and/or, the cycle iteration training times reaches the preset maximum value.
  • the constructed graph convolutional neural network includes a multi-layer graph convolution layer and a classification layer, and preferably includes a three-layer graph convolution layer.
  • the graph convolutional neural network reduces the error between the network output category of a known category of electromagnetic signals and its actual category through a gradient optimization method, and improves the classification accuracy of the network.
  • the gradient optimization method includes Adam, RMSprop, etc.
  • the goal of the graph convolutional neural network is to learn a mapping of signals or features on graph G, and its input layer includes:
  • An N ⁇ D feature matrix X composed of feature descriptions of all nodes. Description of the characteristics of the research network and nodes, that is, the time-frequency characteristics of electromagnetic signals;
  • the network output layer will output the output feature vectors z i ⁇ R F of all nodes (where F represents the number of electromagnetic signal categories), and the output feature vectors of all nodes form an N ⁇ F matrix Z, and then use the Softmax layer to classify the signals , Get the probability distribution vector of each node i belonging to each electromagnetic signal category Estimate the probability distribution vector of the neural network output category through cross-entropy The degree of agreement with the one-hot code y i of the actual category is as follows:
  • the network output is the electromagnetic signal belongs to More accurate category.
  • the electromagnetic signal recognition method based on the implicit knowledge of the graph convolutional network extracts the implicit knowledge of the electromagnetic signal to construct the implicit knowledge base of the electromagnetic signal; based on the implicit knowledge base of the electromagnetic signal, the similarity between different electromagnetic signals is compared, Thereby, an adjacency matrix describing the relationship between electromagnetic signals is generated, and the adjacency matrix is used as the feature description of the graph structure; the time-frequency characteristics of electromagnetic signals are extracted, and then the feature dimensionality is reduced, and the amount of calculation is reduced. Characteristic description; Finally, a graph neural network is built, the adjacency matrix and the time-frequency characteristics of electromagnetic signals are used as the network input, and the category of each electromagnetic signal is output.
  • the electromagnetic signal is characterized as a two-dimensional time-frequency distribution using the Hilbert-Huang transform (HHT), and then the HHT time-frequency distribution is reduced by the local linear embedding method (LLE) to obtain each electromagnetic signal
  • HHT Hilbert-Huang transform
  • LLE local linear embedding method
  • the input layer of the graph convolutional neural network includes an N ⁇ D feature matrix X composed of the time-frequency characteristics of the electromagnetic signal and an adjacency matrix A defined based on the implicit knowledge of the electromagnetic signal. Then superimpose three hidden layers after the input layer to make deeper use of the relationship between nodes.
  • the network output layer will output the output feature vectors z i ⁇ R F of all nodes (where F represents the number of electromagnetic signal categories),
  • the output eigenvectors of all nodes form an N ⁇ F matrix Z, and then use the Softmax layer to classify the signals to obtain the probability distribution vector of each node i belonging to each electromagnetic signal category
  • the degree of consistency with the one-hot code y i of the actual category is reduced by the Adam optimization algorithm to reduce the network output category of the known category of electromagnetic signals Compared with the cross entropy between the actual category y i , when the mean square error of a single cycle is less than the minimum value of 10 -4 , the training ends, and a trained neural network is obtained.
  • the network output is the more accurate electromagnetic signal belonging to category.
  • the electromagnetic signal is characterized as a two-dimensional time-frequency distribution using wavelet transform (WT), and then the WT time-frequency distribution is reduced by linear discriminant analysis (LDA) to obtain the characteristic description x of each electromagnetic signal i i ⁇ R C (where C represents the time-frequency feature number after dimensionality reduction), the feature description of all electromagnetic signals constitutes an N ⁇ C feature matrix X (where N represents the number of all electromagnetic signals).
  • WT wavelet transform
  • LDA linear discriminant analysis
  • the input layer of the graph convolutional neural network includes an N ⁇ D feature matrix X composed of the time-frequency characteristics of the electromagnetic signal and an adjacency matrix A defined based on the implicit knowledge of the electromagnetic signal. Then superimpose three hidden layers after the input layer to make deeper use of the relationship between nodes.
  • the network output layer will output the output feature vectors z i ⁇ R F of all nodes (where F represents the number of electromagnetic signal categories),
  • the output eigenvectors of all nodes form an N ⁇ F matrix Z, and then use the Softmax layer to classify the signals to obtain the probability distribution vector of each node i belonging to each electromagnetic signal category
  • the degree of consistency with the one-hot code y i of the actual category is reduced by the RMSprop optimization algorithm to reduce the network output category of the known category of electromagnetic signals Its actual cross-entropy between categories y i, when the number of cycles reaches the most 10 5 times the value, the end of training, to obtain a trained neural network, network output that is more accurate category belongs to the electromagnetic signal.
  • constructing graph convolutional neural network based on implicit knowledge of electromagnetic signal can mine the relationship between nodes, and for each The characteristics of electromagnetic signal categories are extracted in depth, using more information, and the generalization ability is strong.
  • the adjacency matrix makes full use of the prior knowledge of the signal, which increases the accuracy of the description of the relationship between the signals, improves the classification accuracy, and improves The controllability and interpretability of neural networks are discussed.
  • Fig. 7 is a schematic structural diagram of an electromagnetic signal recognition device for constructing a graph convolutional network based on implicit knowledge according to an embodiment of the present application.
  • the electromagnetic signal recognition device for constructing a graph convolutional network based on implicit knowledge includes: a time-frequency feature extraction module 100, an implicit knowledge graph construction module 200 and an electromagnetic signal classification module 300.
  • the time-frequency feature extraction module 100 is used to obtain multiple electromagnetic signals, extract feature descriptions of multiple electromagnetic signals, and generate a feature matrix as the first type of input information according to the feature descriptions of multiple electromagnetic signals.
  • the implicit knowledge graph construction module 200 is used to mine the implicit knowledge of multiple electromagnetic signals, and construct a graph structure based on the implicit knowledge of electromagnetic signals according to the implicit knowledge as the second type of input information.
  • the electromagnetic signal classification module 300 is used to build a graph convolutional neural network according to the first type of input information and the second type of input information, and use the weakly supervised learning method to perform cyclic and iterative training on the graph convolutional neural network to make the graph convolutional neural network Output multiple types of electromagnetic signals.
  • the electromagnetic signal recognition device can use the characteristics of each node to mine the relationship between the nodes, and perform in-depth extraction of the characteristics of each electromagnetic signal category, use more information, and have a stronger generalization ability.
  • the implicit knowledge includes: modulation type, carrier frequency, symbol rate, radio frequency characteristics, spectrum characteristics and channel characteristics.
  • time-frequency feature extraction module is also used for,
  • Component analysis method linear discriminant analysis method and local linear embedding method are used to reduce the dimension of time-frequency features.
  • the electromagnetic signal classification module is also used to evaluate the accuracy of the categories of the multiple electromagnetic signals output by the graph convolutional neural network by calculating cross entropy.
  • the condition for ending training is: the absolute error of the cross entropy of two consecutive loop iterations training is less than the preset minimum value, And the cross entropy of a single loop training is less than the preset minimum value, and/or, the number of loop iteration training reaches the preset maximum value.
  • constructing graph convolutional neural network based on implicit knowledge of electromagnetic signal can mine the relationship between nodes, and for each The characteristics of electromagnetic signal categories are extracted in depth, using more information, and the generalization ability is strong.
  • the adjacency matrix makes full use of the prior knowledge of the signal, which increases the accuracy of the description of the relationship between the signals, improves the classification accuracy, and improves The controllability and interpretability of neural networks are discussed.
  • first and second are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. Therefore, the features defined with “first” and “second” may explicitly or implicitly include at least one of the features. In the description of feedback in this application, “multiple” means at least two, such as two, three, etc., unless otherwise specifically defined.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Molecular Biology (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Image Analysis (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

本申请反馈公开了一种基于内隐知识构建图卷积网络的电磁信号识别方法及装置，其中，该方法包括：获取多个电磁信号，并提取多个电磁信号的特征描述，根据多个电磁信号的特征描述生成特征矩阵作为第一类输入信息；挖掘多个电磁信号的内隐知识，根据内隐知识构建基于电磁信号内隐知识的图结构作为第二类输入信息；根据第一类输入信息和第二类输入信息搭建图卷积神经网络，利用弱监督学习方法对图卷积神经网络进行循环迭代训练，以使图卷积神经网络输出多个电磁信号的类别。该方法可以利用每个节点的特征，挖掘节点与节点之间的关系，对每个电磁信号类别的特征进行深入的提取，构建图卷积神经网络，对电磁信号进行识别，利用更多的信息，泛化能力更强。

Description

基于内隐知识构建图卷积网络的电磁信号识别方法及装置
相关申请的交叉引用
本申请要求清华大学于2019年05月13日提交的、申请反馈名称为“基于内隐知识构建图卷积网络的电磁信号识别方法及装置”的、中国专利申请号“201910395367.0”的优先权。
技术领域
本申请反馈属于电磁信号智能感知技术领域，特别涉及一种基于内隐知识构建图卷积网络的电磁信号识别方法及装置。
背景技术
当今，随着移动通信设备的不断普及以及物联网技术的蓬勃发展，特别是在5G技术和物联网技术即将大规模应用的背景下，保护无线通讯的安全比如设备验证、非法发射设备监测等，不论是在民用通讯还是工业制造中都变得愈加重要。而信号细微特征的提取在这一领域展现出了巨大的应用价值，它可以利用信号处理技术对无线通信信号进行辨识并提取细微特征，然后与密钥系统配合使用，建立基于软硬件的双重识别体制，使得WLAN系统的信息安全得到提高，因此展现出了极为诱人的前景。而这一机制的核心则在于建立有效的电磁信号识别方法。
传统的识别方法是基于特征匹配、统计决策理论或支持向量机分类等方法。然而需要注意的是这些依赖于人工设计的识别方法在当前越来越复杂的电磁环境下，很难保证其对目标的辨识精度、对识别的响应速度。而通过过去几年机器学习在各领域的发展可以看到，机器学习对于解决感知和识别问题有着更好的表现。因此发展基于机器学习的智能电磁信号识别方法，提高电磁信号识别设备的智能化是十分重要的。
发明内容
本申请反馈旨在至少在一定程度上解决相关技术中的技术问题之一。
为此，本申请反馈的一个目的在于提出一种基于内隐知识构建图卷积网络的电磁信号识别方法，该方法可以利用每个节点的特征，挖掘节点与节点之间的关系，对每个电磁信号类别的特征进行深入的提取，利用更多的信息，泛化能力更强。
本申请反馈的另一个目的在于提出一种基于内隐知识构建图卷积网络的电磁信号识别装置。
为达到上述目的，本申请反馈一方面实施例提出了一种基于内隐知识构建图卷积网络的电磁信号识别方法，包括：
S1，获取多个电磁信号，并提取所述多个电磁信号的特征描述，根据所述多个电磁信号的特征描述生成特征矩阵作为第一类输入信息；
S2，挖掘所述多个电磁信号的内隐知识，根据所述内隐知识构建基于电磁信号内隐知识的图结构作为第二类输入信息；
S3，根据所述第一类输入信息和所述第二类输入信息搭建图卷积神经网络，利用弱监督学习方法对所述图卷积神经网络进行循环迭代训练，以使所述图卷积神经网络输出所述多个电磁信号的类别。
本申请反馈实施例的基于内隐知识构建图卷积网络的电磁信号识别方法，基于电磁信号的内隐知识搭建图卷积神经网络，可以挖掘节点与节点之间的关系，对每个电磁信号类别的特征进行深入的提取，利用更多的信息，泛化能力强，同时充分利用信号的先验知识，增加了信号间的关系描述精确度，提升了分类准确度，提高了神经网络的可控性和可解释性。
另外，根据本申请反馈上述实施例的基于内隐知识构建图卷积网络的电磁信号识别方法还可以具有以下附加的技术特征：
进一步地，所述内隐知识包括：调制类型、载波频率、码元速率、射频特征、频谱特征和信道特征。
进一步地，所述多个电磁信号的特征描述为时频特征；
通过短时傅里叶变换、小波变换和希尔伯特—黄变换提取所述多个电磁信号的特征描述。
进一步地，所述S1还包括：
通过成分分析法、线性判别分析法和局部线性嵌入法对所述时频特征进行降维处理。
进一步地，所述S2，具体包括：
挖掘所述多个电磁信号的内隐知识，以比较不同电磁信号之间的相关性，根据所述不同电磁信号之间的相关性生成邻接矩阵，根据所述邻接矩阵构建所述图结构作为所述第二类输入信息，其中，所述图结构的图节点代表所述不同电磁信号。
进一步地，所述S3，还包括：通过计算交叉熵来评估所述图卷积神经网络输出所述多个电磁信号的类别的准确度。
为达到上述目的，本申请反馈另一方面实施例提出了一种基于内隐知识构建图卷积网 络的电磁信号识别装置，包括：
时频特征提取模块，用于获取多个电磁信号，提取所述多个电磁信号的特征描述，根据所述多个电磁信号的特征描述生成特征矩阵作为第一类输入信息；
内隐知识图构建模块，用于挖掘所述多个电磁信号的内隐知识，根据所述内隐知识构建基于电磁信号内隐知识的图结构作为第二类输入信息；
电磁信号分类模块，用于根据所述第一类输入信息和所述第二类输入信息搭建图卷积神经网络，利用弱监督学习方法对所述图卷积神经网络进行循环迭代训练，以使所述图卷积神经网络输出所述多个电磁信号的类别。
本申请反馈实施例的基于内隐知识构建图卷积网络的电磁信号识别装置，基于电磁信号的内隐知识搭建图卷积神经网络，可以挖掘节点与节点之间的关系，对每个电磁信号类别的特征进行深入的提取，利用更多的信息，泛化能力强，同时充分利用信号的先验知识，增加了信号间的关系描述精确度，提升了分类准确度，提高了神经网络的可控性和可解释性。
另外，根据本申请反馈上述实施例的基于内隐知识构建图卷积网络的电磁信号识别装置还可以具有以下附加的技术特征：
进一步地，所述多个电磁信号的特征描述为时频特征；
通过短时傅里叶变换、小波变换和希尔伯特—黄变换提取所述多个电磁信号的特征描述；
所述内隐知识包括：调制类型、载波频率、码元速率、射频特征、频谱特征和信道特征。
进一步地，所述时频特征提取模块，还用于，
通过成分分析法、线性判别分析法和局部线性嵌入法对所述时频特征进行降维处理。
进一步地，所述电磁信号分类模块，还用于通过计算交叉熵来评估所述图卷积神经网络输出所述多个电磁信号的类别的准确度。
本申请反馈附加的方面和优点将在下面的描述中部分给出，部分将从下面的描述中变得明显，或通过本申请反馈的实践了解到。
附图说明
本申请反馈上述的和/或附加的方面和优点从下面结合附图对实施例的描述中将变得明显和容易理解，其中：
图1为根据本申请反馈一个实施例的基于内隐知识构建图卷积网络的电磁信号识别方法流程图；
图2为根据本申请反馈一个实施例的不同电磁信号的时频图像示意图；
图3为根据本申请反馈一个实施例的电磁信号传播示意图；
图4为根据本申请反馈一个实施例的基于电磁信号内隐知识的图结构示意图；
图5为根据本申请反馈一个实施例的基于电磁信号内隐知识的图卷积神经网络的结构图；
图6为根据本申请反馈又一个实施例的基于内隐知识构建图卷积网络的电磁信号识别方法流程图；
图7为根据本申请反馈一个实施例的基于内隐知识构建图卷积网络的电磁信号识别装置结构示意图。
具体实施方式
下面详细描述本申请反馈的实施例，所述实施例的示例在附图中示出，其中自始至终相同或类似的标号表示相同或类似的元件或具有相同或类似功能的元件。下面通过参考附图描述的实施例是示例性的，旨在用于解释本申请反馈，而不能理解为对本申请反馈的限制。
下面参照附图描述根据本申请反馈实施例提出的基于内隐知识构建图卷积网络的电磁信号识别方法及装置。
首先将参照附图描述根据本申请反馈实施例提出的基于内隐知识构建图卷积网络的电磁信号识别方法。
图1为根据本申请反馈一个实施例的基于内隐知识构建图卷积网络的电磁信号识别方法流程图。
如图1所示，该基于内隐知识构建图卷积网络的电磁信号识别方法包括以下步骤：
步骤S1，获取多个电磁信号，并提取多个电磁信号的特征描述，根据多个电磁信号的特征描述生成特征矩阵作为第一类输入信息。
具体地，通过多种方式提取出多个电磁信号的特征描述构成特征矩阵作为构建的图卷积神经网络的第一类输入信息。
进一步地，电磁信号的特征描述可以有很多种，作为一种方式，提取的多个电磁信号的特征描述为时频特征，提取特征描述的方式又多种，包括但不限于短时傅里叶变换(Short Time Fourier Transform，STFT)、小波变换(Wavelet Transform，WT)和希尔伯特—黄变换(Hilbert-Huang Transform，HHT)。
进一步地，S1还包括：
通过成分分析法(Principal Component Analysis，PCA)、线性判别分析法(Linear  Discriminant Analysis，LDA)和局部线性嵌入法(Locally linear embedding，LLE)对时频特征进行降维处理。
具体地，由于电磁信号时频特征具有更高的辨识性，因此将电磁信号利用短时傅里叶变换(Short-time Fourier Transform，STFT)表征为二维时频分布即可作为节点的特征描述，如图2所示，给出了四种不同电磁信号的STFT时频图像，可见不同种类电磁信号的时频分布有明显区别。然而，兼顾到分辨精度，时频图像的维度通常较高，会导致巨大的计算量，因此，需要对STFT时频分布利用主成分分析法进行降维处理，从而得到每个电磁信号i的特征描述x i∈R C(其中C代表降维后的时频特征数)，所有电磁信号的特征描述构成N×C的特征矩阵X(其中N代表所有电磁信号的数目)。
步骤S2，挖掘多个电磁信号的内隐知识，根据内隐知识构建基于电磁信号内隐知识的图结构作为第二类输入信息。
进一步地，内隐知识包括但不限于调制类型、载波频率、码元速率、射频特征、频谱特征和信道特征。
进一步地，S2具体包括：挖掘多个电磁信号的内隐知识，以比较不同电磁信号之间的相关性，根据不同电磁信号之间的相关性生成邻接矩阵，根据邻接矩阵构建图结构作为第二类输入信息，其中，图结构的图节点代表不同电磁信号。
可以理解的是，构建的图结构中，其节点包含的信息为不同电磁信号的特征描述，节点与节点之间的连接关系可由邻接矩阵给出。
具体的，提取已知类别的电磁信号集合X 1和未知类别的电磁信号集合X 2中所有电磁信号的内隐知识，包括其调制类型、载波频率、码元速率、信道传播方式等，这些内隐知识包含在信号的传输过程中，如图3所示，构成了信号的内隐知识集合K＝{k 1,k 2,…,k L}，其中L为内隐知识的种类数。
如图3所示，图3(a)、图3(b)、图3(c)、图3(d)、图3(e)、图3(f)分别代表信号、调制、数模转换、信道、射频、天线，为对图3(g)的进一步说明。
具体地，根据内隐知识构建图结构，我们构建的基于电磁信号内隐知识的图结构，如图4所示，图G是由顶点集V和连接顶点的边集E构成的离散结构，因此可以表示成G＝(V,E)。对于研究的基于电磁信号内隐知识的图结构，其每个顶点为已知类别电磁信号i∈X 1或未知类别电磁信号i∈X 2的特征描述x i，邻接矩阵A∈R N×N则根据电磁信号的内隐知识集合K来构建：对于邻接矩阵中任意位置的元素A ij，若所对应的两个电磁信号i和j都类别已知，那么如果属于同一类，则A ij＝1；如果属于不同类，则A ij＝0；若所对应的两个电磁信号中存在类别未知的信号，则需比较二者的内隐知识集合，计算相关性作为元素A ij 的值。
可以理解的是，邻接矩阵的值介于0到1之间，是由电磁信号的内隐知识决定的，任意两个电磁信号的内隐知识相关性越高，则邻接矩阵对应元素的值越大。
步骤S3，根据第一类输入信息和第二类输入信息搭建图卷积神经网络，利用弱监督学习方法对图卷积神经网络进行循环迭代训练，以使图卷积神经网络输出多个电磁信号的类别。
进一步地，S3还包括：通过计算交叉熵来评估图卷积神经网络输出多个电磁信号的类别与其实际类别的一致程度。
通过多次的循环迭代训练，提高图卷积神经网络输出多个电磁信号的类别的准确度，其中，训练中止的条件为：连续两次循环循迭代训练的交叉熵的绝对误差小于预设最小值，且单次循环训练的交叉熵小于预设最小值，和\或，循环迭代训练次数达到预设最大值。
具体地，搭建的图卷积神经网络包含多层图卷积层和分类层，优选地包括三层图卷积层。
图卷积神经网络通过梯度优化方法降低已知类别电磁信号的网络输出类别与其实际类别之间的误差，提高网络的分类精度，优选地梯度优化方法包括Adam、RMSprop等。
具体地，图卷积神经网络的目标是要学习图G上的信号或特征的一个映射，其输入层包括：
(1)所有节点的特征描述构成的N×D的特征矩阵X。对于研究的网络、节点特征描述即电磁信号的时频特征；
(2)矩阵形式的图结构的特征描述。对于我们研究的网络，图结构特征描述即基于电磁信号内隐知识定义的邻接矩阵A。
通过在输入层之后叠加三层隐藏层，如图5所示，更深度地利用节点与节点之间的关系，从而归纳出每个类别更泛化的特征，提高分类精度。网络输出层会输出所有节点的输出特征向量z i∈R F(其中F表示电磁信号类别的数目)，所有节点的输出特征向量构成N×F的矩阵Z，再利用Softmax层，对信号进行分类，得到每个节点i属于每个电磁信号类别的概率分布向量
Figure PCTCN2019098100-appb-000001
通过交叉熵(cross-entropy)来估量神经网络输出类别的概率分布向量
Figure PCTCN2019098100-appb-000002
与其实际类别的one-hot编码y i的一致程度，公式如下:
Figure PCTCN2019098100-appb-000003
通过Adam优化算法降低已知类别电磁信号的网络输出类别
Figure PCTCN2019098100-appb-000004
与其实际类别y i之间的交叉熵，当连续两次循环的均方误差值的绝对误差小于10 -5时，训练结束，从而得到了训练好的神经网络，网络输出即为电磁信号所属的较准确的类别。
综上，基于内隐知识构建图卷积网络的电磁信号识别方法通过提取电磁信号内隐知识，构建电磁信号内隐知识库；基于电磁信号内隐知识库比较不同电磁信号之间的相似性，从而生成描述电磁信号之间关系的邻接矩阵，将邻接矩阵作为图结构的特征描述；提取电磁信号时频特征，然后进行特征降维，降低运算量，将降维后的时频特征作为图节点的特征描述；最后搭建图神经网络，将邻接矩阵和电磁信号的时频特征作为网络输入，输出每个电磁信号所属类别。
下面通过具体实施例对基于内隐知识构建图卷积网络的电磁信号识别方法详细描述。
如图6所示，通过希尔伯特—黄变换提取多个电磁信号的特征描述。
1)提取电磁信号的特征描述；
具体的，将电磁信号利用希尔伯特—黄变换(HHT)表征为二维时频分布，然后对HHT时频分布利用局部线性嵌入法(LLE)进行降维处理，从而得到每个电磁信号i的特征描述x i∈R C(其中C代表降维后的时频特征数)，所有电磁信号的特征描述构成N×C的特征矩阵X(其中N代表所有电磁信号的数目)。
2)提取电磁信号内隐知识；
具体的，提取已知类别的电磁信号集合X 1和未知类别的电磁信号集合X 2中所有电磁信号的内隐知识，包括其调制类型、载波频率、码元速率、信道传播方式等，它们构成了信号的内隐知识集合K＝{k 1,k 2,…,k L}，其中L为内隐知识的种类数。
3)构建基于电磁信号内隐知识的图G；
具体的，图G的每个顶点为已知类别电磁信号i∈X 1或未知类别电磁信号i∈X 2的特征描述x i，邻接矩阵A∈R N×N则根据电磁信号的内隐知识集合K来构建：对于邻接矩阵中任意位置的元素A ij，若所对应的两个电磁信号i和j都类别已知，那么如果属于同一类，则A ij＝1；如果属于不同类，则A ij＝0；若所对应的两个电磁信号中存在类别未知的信号，则需比较二者的内隐知识集合，计算相关性作为元素A ij的值。
4)搭建基于电磁信号内隐知识的图卷积神经网络。
具体的，图卷积神经网络的输入层包括电磁信号的时频特征构成的N×D的特征矩阵X和基于电磁信号内隐知识定义的邻接矩阵A。然后在输入层之后叠加三层隐藏层，更深度地利用节点与节点之间的关系，网络输出层会输出所有节点的输出特征向量z i∈R F(其中F表示电磁信号类别的数目)，所有节点的输出特征向量构成N×F的矩阵Z，再利用Softmax层，对信号进行分类，得到每个节点i属于每个电磁信号类别的概率分布向量
Figure PCTCN2019098100-appb-000005
通过交叉熵来估量神经网络输出类别的概率分布向量
Figure PCTCN2019098100-appb-000006
与其实际类别的one-hot编码y i的一致程度，通过Adam优化算法降低已知类别电磁信号的网络输出类别
Figure PCTCN2019098100-appb-000007
与其实际类别y i之间的交叉熵， 当单次循环的均方误差小于最小值10 -4时，训练结束，从而得到了训练好的神经网络，网络输出即为电磁信号所属的较准确的类别。
进一步地，通过小波变换对多个电磁信号的特征描述进行提取。如图6所示。
1)提取电磁信号的特征描述；
具体的，将电磁信号利用小波变换(WT)表征为二维时频分布，然后对WT时频分布利用线性判别分析法(LDA)进行降维处理，从而得到每个电磁信号i的特征描述x i∈R C(其中C代表降维后的时频特征数)，所有电磁信号的特征描述构成N×C的特征矩阵X(其中N代表所有电磁信号的数目)。
2)提取电磁信号内隐知识；
具体的，提取已知类别的电磁信号集合X 1和未知类别的电磁信号集合X 2中所有电磁信号的内隐知识，包括其调制类型、载波频率、码元速率、信道传播方式等，它们构成了信号的内隐知识集合K＝{k 1,k 2,…,k L}，其中L为内隐知识的种类数。
3)构建基于电磁信号内隐知识的图G；
具体的，图G的每个顶点为已知类别电磁信号i∈X 1或未知类别电磁信号i∈X 2的特征描述x i，邻接矩阵A∈R N×N则根据电磁信号的内隐知识集合K来构建：对于邻接矩阵中任意位置的元素A ij，若所对应的两个电磁信号i和j都类别已知，那么如果属于同一类，则A ij＝1；如果属于不同类，则A ij＝0；若所对应的两个电磁信号中存在类别未知的信号，则需比较二者的内隐知识集合，计算相关性作为元素A ij的值。
4)搭建基于电磁信号内隐知识的图卷积神经网络；
具体的，图卷积神经网络的输入层包括电磁信号的时频特征构成的N×D的特征矩阵X和基于电磁信号内隐知识定义的邻接矩阵A。然后在输入层之后叠加三层隐藏层，更深度地利用节点与节点之间的关系，网络输出层会输出所有节点的输出特征向量z i∈R F(其中F表示电磁信号类别的数目)，所有节点的输出特征向量构成N×F的矩阵Z，再利用Softmax层，对信号进行分类，得到每个节点i属于每个电磁信号类别的概率分布向量
Figure PCTCN2019098100-appb-000008
通过交叉熵来估量神经网络输出类别的概率分布向量
Figure PCTCN2019098100-appb-000009
与其实际类别的one-hot编码y i的一致程度，通过RMSprop优化算法降低已知类别电磁信号的网络输出类别
Figure PCTCN2019098100-appb-000010
与其实际类别y i之间的交叉熵，当循环次数达到最值10 5次时，训练结束，从而得到了训练好的神经网络，网络输出即为电磁信号所属的较准确的类别。
根据本申请反馈实施例提出的基于内隐知识构建图卷积网络的电磁信号识别方法，基于电磁信号的内隐知识搭建图卷积神经网络，可以挖掘节点与节点之间的关系，对每个电磁信号类别的特征进行深入的提取，利用更多的信息，泛化能力强，同时通过邻接矩阵充 分利用信号的先验知识，增加了信号间的关系描述精确度，提升了分类准确度，提高了神经网络的可控性和可解释性。
其次参照附图描述根据本申请反馈实施例提出的基于内隐知识构建图卷积网络的电磁信号识别装置。
图7为根据本申请反馈一个实施例的基于内隐知识构建图卷积网络的电磁信号识别装置结构示意图。
如图7所示，该基于内隐知识构建图卷积网络的电磁信号识别装置包括：时频特征提取模块100、内隐知识图构建模块200和电磁信号分类模块300。
其中，时频特征提取模块100，用于获取多个电磁信号，提取多个电磁信号的特征描述，根据多个电磁信号的特征描述生成特征矩阵作为第一类输入信息。
内隐知识图构建模块200，用于挖掘多个电磁信号的内隐知识，根据内隐知识构建基于电磁信号内隐知识的图结构作为第二类输入信息。
电磁信号分类模块300，用于根据第一类输入信息和第二类输入信息搭建图卷积神经网络，利用弱监督学习方法对图卷积神经网络进行循环迭代训练，以使图卷积神经网络输出多个电磁信号的类别。
该电磁信号识别装置可以利用每个节点的特征，挖掘节点与节点之间的关系，对每个电磁信号类别的特征进行深入的提取，利用更多的信息，泛化能力更强。
进一步地，多个电磁信号的特征描述为时频特征；
通过短时傅里叶变换、小波变换和希尔伯特—黄变换提取多个电磁信号的特征描述；
内隐知识包括：调制类型、载波频率、码元速率、射频特征、频谱特征和信道特征。
进一步地，时频特征提取模块，还用于，
通过成分分析法、线性判别分析法和局部线性嵌入法对时频特征进行降维处理。
进一步地，电磁信号分类模块，还用于通过计算交叉熵来评估图卷积神经网络输出多个电磁信号的类别的准确度。
通过多次迭代训练使图卷积神经网络输出的多个电磁信号的类别的准确度更高，结束训练的条件为：连续两次循环迭代循训练的交叉熵的绝对误差小于预设最小值，且单次循环训练的交叉熵小于预设最小值，和\或，循环迭代训练次数达到预设最大值。
需要说明的是，前述对基于内隐知识构建图卷积网络的电磁信号识别方法实施例的解释说明也适用于该实施例的装置，此处不再赘述。
根据本申请反馈实施例提出的基于内隐知识构建图卷积网络的电磁信号识别装置，基于电磁信号的内隐知识搭建图卷积神经网络，可以挖掘节点与节点之间的关系，对每个电磁信号类别的特征进行深入的提取，利用更多的信息，泛化能力强，同时通过邻接矩阵充 分利用信号的先验知识，增加了信号间的关系描述精确度，提升了分类准确度，提高了神经网络的可控性和可解释性。
此外，术语“第一”、“第二”仅用于描述目的，而不能理解为指示或暗示相对重要性或者隐含指明所指示的技术特征的数量。由此，限定有“第一”、“第二”的特征可以明示或者隐含地包括至少一个该特征。在本申请反馈的描述中，“多个”的含义是至少两个，例如两个，三个等，除非另有明确具体的限定。
在本说明书的描述中，参考术语“一个实施例”、“一些实施例”、“示例”、“具体示例”、或“一些示例”等的描述意指结合该实施例或示例描述的具体特征、结构、材料或者特点包含于本申请反馈的至少一个实施例或示例中。在本说明书中，对上述术语的示意性表述不必须针对的是相同的实施例或示例。而且，描述的具体特征、结构、材料或者特点可以在任一个或多个实施例或示例中以合适的方式结合。此外，在不相互矛盾的情况下，本领域的技术人员可以将本说明书中描述的不同实施例或示例以及不同实施例或示例的特征进行结合和组合。
尽管上面已经示出和描述了本申请反馈的实施例，可以理解的是，上述实施例是示例性的，不能理解为对本申请反馈的限制，本领域的普通技术人员在本申请反馈的范围内可以对上述实施例进行变化、修改、替换和变型。

Claims (10)

  1. 一种基于内隐知识构建图卷积网络的电磁信号识别方法，其特征在于，包括以下步骤：
    S1，获取多个电磁信号，并提取所述多个电磁信号的特征描述，根据所述多个电磁信号的特征描述生成特征矩阵作为第一类输入信息；
    S2，挖掘所述多个电磁信号的内隐知识，根据所述内隐知识构建基于电磁信号内隐知识的图结构作为第二类输入信息；
    S3，根据所述第一类输入信息和所述第二类输入信息搭建图卷积神经网络，利用弱监督学习方法对所述图卷积神经网络进行循环迭代训练，以使所述图卷积神经网络输出所述多个电磁信号的类别。
  2. 根据权利要求1所述的方法，其特征在于，所述内隐知识包括：调制类型、载波频率、码元速率、射频特征、频谱特征和信道特征。
  3. 根据权利要求1所述的方法，其特征在于，
    所述多个电磁信号的特征描述为时频特征；
    通过短时傅里叶变换、小波变换和希尔伯特—黄变换提取所述多个电磁信号的特征描述。
  4. 根据权利要求3所述的方法，其特征在于，所述S1还包括：
    通过成分分析法、线性判别分析法和局部线性嵌入法对所述时频特征进行降维处理。
  5. 根据权利要求1所述的方法，其特征在于，所述S2，具体包括：
    挖掘所述多个电磁信号的内隐知识，以比较不同电磁信号之间的相关性，根据所述不同电磁信号之间的相关性生成邻接矩阵，根据所述邻接矩阵构建所述图结构作为所述第二类输入信息，其中，所述图结构的图节点代表所述不同电磁信号。
  6. 根据权利要求1所述的方法，其特征在于，所述S3还包括：通过计算交叉熵来评估所述图卷积神经网络输出所述多个电磁信号的类别的准确度。
  7. 一种基于内隐知识构建图卷积网络的电磁信号识别装置，其特征在于，包括：
    时频特征提取模块，用于获取多个电磁信号，提取所述多个电磁信号的特征描述，根据所述多个电磁信号的特征描述生成特征矩阵作为第一类输入信息；
    内隐知识图构建模块，用于挖掘所述多个电磁信号的内隐知识，根据所述内隐知识构建基于电磁信号内隐知识的图结构作为第二类输入信息；
    电磁信号分类模块，用于根据所述第一类输入信息和所述第二类输入信息搭建图卷积神经网络，利用弱监督学习方法对所述图卷积神经网络进行循环迭代训练，以使所述图卷 积神经网络输出所述多个电磁信号的类别。
  8. 根据权利要求7所述的装置，其特征在于，
    所述多个电磁信号的特征描述为时频特征；
    通过短时傅里叶变换、小波变换和希尔伯特—黄变换提取所述多个电磁信号的特征描述；
    所述内隐知识包括：调制类型、载波频率、码元速率、射频特征、频谱特征和信道特征。
  9. 根据权利要求7所述的装置，其特征在于，所述时频特征提取模块，还用于，
    通过成分分析法、线性判别分析法和局部线性嵌入法对所述时频特征进行降维处理。
  10. 根据权利要求7所述的装置，其特征在于，所述电磁信号分类模块，还用于通过计算交叉熵来评估所述图卷积神经网络输出所述多个电磁信号的类别的准确度。
PCT/CN2019/098100 2019-05-13 2019-07-29 基于内隐知识构建图卷积网络的电磁信号识别方法及装置 Ceased WO2020228141A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910395367.0A CN110210330B (zh) 2019-05-13 2019-05-13 基于内隐知识构建图卷积网络的电磁信号识别方法及装置
CN201910395367.0 2019-05-13

Publications (1)

Publication Number Publication Date
WO2020228141A1 true WO2020228141A1 (zh) 2020-11-19

Family

ID=67787201

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/098100 Ceased WO2020228141A1 (zh) 2019-05-13 2019-07-29 基于内隐知识构建图卷积网络的电磁信号识别方法及装置

Country Status (2)

Country Link
CN (1) CN110210330B (zh)
WO (1) WO2020228141A1 (zh)

Cited By (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113095137A (zh) * 2021-03-10 2021-07-09 南京大学 一种基于机器学习的信号特征识别装置和方法
CN113283488A (zh) * 2021-05-18 2021-08-20 清华大学 一种基于学习行为的认知诊断方法及系统
CN113313039A (zh) * 2021-05-31 2021-08-27 西安电子科技大学 基于动作知识库与集成学习的视频行为识别方法及系统
CN113822291A (zh) * 2021-06-15 2021-12-21 腾讯科技（深圳）有限公司 一种图像处理方法、装置、设备及存储介质
CN114186593A (zh) * 2021-12-14 2022-03-15 深圳潜翔定制设计有限公司 噪声条件下轴承故障识别方法、装置、设备和存储介质
CN114741519A (zh) * 2022-02-18 2022-07-12 北京邮电大学 一种基于图卷积神经网络和知识库的论文相关性分析方法
CN116340533A (zh) * 2023-02-23 2023-06-27 中国人民解放军军事科学院系统工程研究院 一种基于知识图谱的星载电磁频谱大数据智能处理系统
CN116467626A (zh) * 2023-04-11 2023-07-21 四川九洲电器集团有限责任公司 一种辐射源个体识别方法
CN117155746A (zh) * 2023-10-31 2023-12-01 中孚安全技术有限公司 电磁信号的合并处理方法、系统及介质
CN117171561A (zh) * 2023-08-23 2023-12-05 中国工程物理研究院电子工程研究所 一种基于数据驱动和知识约束的信号载波检测与识别方法
CN117807529A (zh) * 2024-02-29 2024-04-02 南京工业大学 一种信号发生器输出信号的调制方式识别方法及系统
CN118018304A (zh) * 2024-03-05 2024-05-10 桂林电子科技大学 基于知识图谱和残差图卷积网络的内部威胁检测方法
CN118298252A (zh) * 2024-06-06 2024-07-05 中国测绘科学研究院 一种基于图表征学习的城市建筑群功能分类方法及设备
CN119202852A (zh) * 2024-11-14 2024-12-27 杭州电子科技大学 一种基于特征嵌入的图神经网络自动调制分类方法

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111612051B (zh) * 2020-04-30 2023-06-20 杭州电子科技大学 一种基于图卷积神经网络的弱监督目标检测方法
CN113759206B (zh) * 2020-06-05 2026-04-14 中国电力科学研究院有限公司 一种配电网故障类型的判断方法和系统
CN114417938B (zh) * 2022-01-27 2023-06-27 中国人民解放军32802部队 一种利用知识向量嵌入的电磁目标分类方法
CN114911776A (zh) * 2022-06-21 2022-08-16 平安银行股份有限公司 主题模型生成方法、计算机设备和计算机可读存储介质
CN115422971B (zh) * 2022-08-30 2024-11-05 华中科技大学 基于图网络学习的盾构开挖面稳定性智能预测方法及设备
CN116165703B (zh) * 2023-02-24 2025-06-03 西安石油大学 基于流形学习算法的地震数据特征提取方法及系统

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106446777A (zh) * 2016-08-29 2017-02-22 广东工业大学 基于卷积神经网络的无创血糖数据处理方法及系统
CN107024685A (zh) * 2017-04-10 2017-08-08 北京航空航天大学 一种基于距离‑速度特征的手势识别方法
CN108021871A (zh) * 2017-11-22 2018-05-11 华南理工大学 一种基于主成分分析的特征频率提取方法

Family Cites Families (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105718944A (zh) * 2016-01-19 2016-06-29 上海交通大学 基于核空间的深度散射卷积网络学习方法和系统
CN108234370B (zh) * 2017-12-22 2020-12-15 西安电子科技大学 基于卷积神经网络的通信信号调制方式识别方法
CN108304795B (zh) * 2018-01-29 2020-05-12 清华大学 基于深度强化学习的人体骨架行为识别方法及装置
CN108549863B (zh) * 2018-04-11 2019-11-26 腾讯科技（深圳）有限公司 人体姿态预测方法、装置、设备及存储介质
CN108648095A (zh) * 2018-05-10 2018-10-12 浙江工业大学 一种基于图卷积网络梯度的节点信息隐藏方法
CN108764077B (zh) * 2018-05-15 2021-03-19 重庆邮电大学 一种基于卷积神经网络的数字信号调制分类方法
CN108874914B (zh) * 2018-05-29 2021-11-02 吉林大学 一种基于图卷积与神经协同过滤的信息推荐方法
CN109033738B (zh) * 2018-07-09 2022-01-11 湖南大学 一种基于深度学习的药物活性预测方法
CN109255791A (zh) * 2018-07-19 2019-01-22 杭州电子科技大学 一种基于图卷积神经网络的形状协同分割方法
CN109492691A (zh) * 2018-11-07 2019-03-19 南京信息工程大学 一种超图卷积网络模型及其半监督分类方法
CN109636061B (zh) * 2018-12-25 2023-04-18 深圳市南山区人民医院 医保欺诈预测网络的训练方法、装置、设备及存储介质
CN109583519A (zh) * 2018-12-27 2019-04-05 中国石油大学(华东) 一种基于p-Laplacian图卷积神经网络的半监督分类方法
CN110427835B (zh) * 2019-07-11 2022-03-11 清华大学 图卷积网络与迁移学习的电磁信号识别方法和装置

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106446777A (zh) * 2016-08-29 2017-02-22 广东工业大学 基于卷积神经网络的无创血糖数据处理方法及系统
CN107024685A (zh) * 2017-04-10 2017-08-08 北京航空航天大学 一种基于距离‑速度特征的手势识别方法
CN108021871A (zh) * 2017-11-22 2018-05-11 华南理工大学 一种基于主成分分析的特征频率提取方法

Cited By (19)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113095137A (zh) * 2021-03-10 2021-07-09 南京大学 一种基于机器学习的信号特征识别装置和方法
CN113095137B (zh) * 2021-03-10 2023-12-29 南京大学 一种基于机器学习的信号特征识别装置和方法
CN113283488A (zh) * 2021-05-18 2021-08-20 清华大学 一种基于学习行为的认知诊断方法及系统
CN113313039B (zh) * 2021-05-31 2023-07-25 西安电子科技大学 基于动作知识库与集成学习的视频行为识别方法及系统
CN113313039A (zh) * 2021-05-31 2021-08-27 西安电子科技大学 基于动作知识库与集成学习的视频行为识别方法及系统
CN113822291A (zh) * 2021-06-15 2021-12-21 腾讯科技（深圳）有限公司 一种图像处理方法、装置、设备及存储介质
CN114186593A (zh) * 2021-12-14 2022-03-15 深圳潜翔定制设计有限公司 噪声条件下轴承故障识别方法、装置、设备和存储介质
CN114741519A (zh) * 2022-02-18 2022-07-12 北京邮电大学 一种基于图卷积神经网络和知识库的论文相关性分析方法
CN116340533A (zh) * 2023-02-23 2023-06-27 中国人民解放军军事科学院系统工程研究院 一种基于知识图谱的星载电磁频谱大数据智能处理系统
CN116340533B (zh) * 2023-02-23 2023-09-01 中国人民解放军军事科学院系统工程研究院 一种基于知识图谱的星载电磁频谱大数据智能处理系统
CN116467626A (zh) * 2023-04-11 2023-07-21 四川九洲电器集团有限责任公司 一种辐射源个体识别方法
CN117171561A (zh) * 2023-08-23 2023-12-05 中国工程物理研究院电子工程研究所 一种基于数据驱动和知识约束的信号载波检测与识别方法
CN117155746A (zh) * 2023-10-31 2023-12-01 中孚安全技术有限公司 电磁信号的合并处理方法、系统及介质
CN117155746B (zh) * 2023-10-31 2024-02-23 中孚安全技术有限公司 电磁信号的合并处理方法、系统及介质
CN117807529A (zh) * 2024-02-29 2024-04-02 南京工业大学 一种信号发生器输出信号的调制方式识别方法及系统
CN117807529B (zh) * 2024-02-29 2024-05-07 南京工业大学 一种信号发生器输出信号的调制方式识别方法及系统
CN118018304A (zh) * 2024-03-05 2024-05-10 桂林电子科技大学 基于知识图谱和残差图卷积网络的内部威胁检测方法
CN118298252A (zh) * 2024-06-06 2024-07-05 中国测绘科学研究院 一种基于图表征学习的城市建筑群功能分类方法及设备
CN119202852A (zh) * 2024-11-14 2024-12-27 杭州电子科技大学 一种基于特征嵌入的图神经网络自动调制分类方法

Also Published As

Publication number Publication date
CN110210330A (zh) 2019-09-06
CN110210330B (zh) 2021-05-28

Similar Documents

Publication Publication Date Title
CN110210330B (zh) 基于内隐知识构建图卷积网络的电磁信号识别方法及装置
CN112508085B (zh) 基于感知神经网络的社交网络链路预测方法
CN111582320B (zh) 一种基于半监督学习的动态个体识别方法
CN109063724B (zh) 一种增强型生成式对抗网络以及目标样本识别方法
CN113313156A (zh) 一种基于时序负载流量指纹的物联网设备识别方法及系统
CN115907001A (zh) 基于知识蒸馏的联邦图学习方法及自动驾驶方法
CN115114958B (zh) 一种基于监督对比学习的电磁信号开集识别方法
CN104462196A (zh) 多特征联合哈希信息检索方法
CN114611601A (zh) 一种基于孪生网络的小样本辐射源个体识别方法
Sheng et al. Image splicing detection based on Markov features in discrete octonion cosine transform domain
CN117671666A (zh) 一种基于自适应图卷积神经网络的目标识别方法
Yong‐dong et al. Kernel PCA for road traffic data non‐linear feature extraction
CN111046896B (zh) 一种跳频信号电台分选方法
CN113360725B (zh) 基于边缘协同分类的电力时序数据检索方法
Qiu et al. Semi-supervised specific emitter identification based on contrastive learning and data augmentation
Jiang et al. Towards semi-supervised classification of abnormal spectrum signals based on deep learning
CN111652021B (zh) 一种基于bp神经网络的人脸识别方法及系统
Feng et al. Fine-grained transductive prototypical network based few-shot signal modulation classification using coarse labels
Yang et al. FedMMD: Heterogenous federated learning based on multi-teacher and multi-feature distillation
CN112465884B (zh) 基于生成式特征表示网络的多元遥感影像变化检测方法
Sima et al. MSO‐DETR: Metric space optimization for few‐shot object detection
CN117938596B (zh) 一种基于解耦Hilbert-Schmidt信息瓶颈DHSIB的自动调制分类对抗防御方法
Tang et al. Model adaptation via credible local context representation
Luo [Retracted] Analysis of Artistic Modeling of Opera Stage Clothing Based on Big Data Clustering Algorithm
CN112329734B (zh) 基于两模态图强化特征学习的人脸识别方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19928575

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19928575

Country of ref document: EP

Kind code of ref document: A1