Intelligent data and knowledge-driven method for modulation recognition

An intelligent data and knowledge-driven method for a modulation recognition, comprising the following steps:

• step (1):
• collecting spectrum data;
• step (2):
• constructing corresponding attribute vector labels for different modulation schemes;
• step (3):
• constructing and pre-training an attribute learning model based on the attribute vector labels for the different modulation schemes;
• step (4):
• constructing and pre-training a visual model for the modulation recognition;
• step (5):
• constructing a feature space transformation model, and constructing an intelligent data and knowledge-driven model for the modulation recognition based on the attribute learning model and the visual model;
• step (6):
• transferring parameters of a pre-trained visual model and a pre-trained attribute learning model and retraining the feature space transformation model;
• step (7):
• determining whether a training on a network is completed;
• and if yes, performing step (8);
• or if not, increasing a number of training iterations by 1 and retraining the feature space transformation model;
• step (8):
• inputting test set data into the network;
• and
• step (9):
• outputting a classification result;
• wherein the constructing corresponding attribute vector labels for the different modulation schemes in the step (2) specifically comprises:
• using a binary phase shift keying, a quaternary phase shift keying, a 16 quadrature amplitude modulation and a 64 quadrature amplitude modulation as recognition objects to construct a six-dimensional attribute feature vector,
• wherein each dimension indicates a different meaning, a first dimension indicates whether a base number of a classification target is binary, a second dimension indicates whether the base number of the classification target is quaternary, a third dimension indicates whether the base number of the classification target is hexadecimal, a fourth dimension indicates whether the base number of the classification target is base 64, a fifth dimension indicates whether the classification target is a phase shift keying, and a sixth dimension indicates whether the classification target is a quadrature amplitude modulation;
• wherein the constructing and pre-training the attribute learning model based on the attribute vector labels for the different modulation schemes in the step (3) specifically comprises:
• obtaining the attribute learning model by modifying an existing residual neural network,
• wherein a residual unit consists of two convolutional layers with a kernel size of 3×3;
• a batch normalization layer standardizes intermediate data in a middle layer of the network, a linear correction unit is connected after each batch normalization layer as an activation function to introduce a nonlinearity into the network, a residual stack unit is constructed by sequentially connecting one 1×1 convolutional layer, two residual units, and one maximum pooling layer, the convolutional layer and the residual units are used for a feature extraction, and the maximum pooling layer is used to compress features to reduce feature dimensionalities;
• and
• when the attribute learning model is trained, trainable parameters of the network are randomly initialized:
• an initialized number of network training epochs is 1, a maximum number of epochs is 50, a learning rate is 0.001, a stochastic gradient descent (SGD) optimization algorithm is configured as a network training optimizer, and a mean squared error (MSE) loss function is selected to calculate a difference between network output and a real attribute label;
• wherein the constructing and pre-training the visual model for the modulation recognition specifically comprises:
• forming a visual training model by a multi-scale module, a global average pooling layer, a fully connected layer and a classification layer,
• wherein the multi-scale module first uses a convolutional layer with a kernel size of 3×1 and a stride of 2 to reduce the feature dimensionalities, then uses a plurality of convolutional layers with different kernel sizes to learn multidimensional features, and finally splices the multidimensional features;
• and
• a fully connected layer with the linear correction unit as the activation function is used after the global average pooling layer to reduce the feature dimensionalities, and an expression of the activation function is as follows:
• ƒ(z)=max(0,z)
• max(•) indicates calculating a maximum value of an object in parentheses, z indicates input data of the activation function;
• and
• when the visual model is trained, the trainable parameters of the network are randomly initialized:
• the initialized number of the network training epochs is 1, a maximum number of iterations is 20, the learning rate is 0.001, an Adam optimization algorithm is configured as the network training optimizer, and a cross-entropy loss function is selected to calculate the difference between the network output and the real attribute label;
• wherein the constructing the feature space transformation model, and constructing the intelligent data and knowledge-driven model for the modulation recognition based on the attribute learning model and the visual model in the step (5) specifically comprises:
• obtaining a first branch as a visual coding branch by removing a last classification layer of the visual model, specifically, inputting I/Q raw data Ii to the visual coding branch, extracting, by the multi-scale module, the multidimensional features of the input data, flattening, by the global average pooling layer, the multidimensional features, and outputting, by the fully connected layer, a D-dimensional feature vector ϕ1(Ii)∈D×1;
• performing, by a second branch consisting of the attribute learning model, an attribute semantic embedding, wherein inputting the raw data Ii, and outputting an L-dimensional attribute feature vector ϕ2(Ii)∈L×1;
• and
• transferring attribute features to a visual feature space by the feature space transformation model, wherein the feature space transformation model consists of two fully connected layers with linear correction units as activation functions, the L-dimensional attribute feature vector ϕ2(Ii) is configured as an input, the feature space transformation model outputs an embedding vector ϕ3(ϕ2(Ii))∈D×1 of a same dimension as a visual feature, and an output of the feature space transformation model is as follows:
• ϕ3(ϕ2(Ii))=ƒ2(W2ƒ1(W1ϕ2(Ii)))
• W1∈L×M indicates a weight matrix of a first fully connected layer, W2∈M×D indicates a weight matrix of a second fully connected layer, M indicates an output dimension of the first fully connected layer, which is equal to an input dimension of the second fully connected layer, and ƒ1(•) and ƒ2(•) respectively indicate the activation functions connected behind the two fully connected layers and are used to introduce the nonlinearity to a network structure.