Research On Modulation Classification Method Of Wireless Signal Based On Semi-Supervised Learning

The autoencoder consists of an encoder and a decoder.The encoder converts the input data into intermediate-level feature representations,and the decoder uses the intermediate-level feature representations to reconstruct the input data.As a powerful feature extractor,the autoencoder can be applied to Automatic Modulation Classification(AMC).AMC can automatically identify the modulation of wireless signals,and plays a key role in cognitive radio and radio spectrum monitoring as well as in other civilian and military fields.The deep neural networks composed of autoencoders can extract the abstract features of the input data through unsupervised learning.However,while extracting abstract features,the deep neural network will discard a large number of detailed features some of which can effectively improve the classification accuracy.For this reason,this paper proposes a feature extraction method based on stacked convolutional denoising autoencoders(SCDAE).The improved SCDAE adds a connection channel between its low-level network and high-level network,making the detailed features extracted from lower layer networks can be effectively combined with abstract features extracted from high-level networks.When using the improved SCDAE for wireless signal feature extraction,noise-damaged wireless signal data is exploited to train the network.The training goal is to make the SCDAE decoder output is as close as possible to the original input data.After the network training is completed,the wireless signal data is input into the SCDAE encoder,and the output of the encoder is the extracted signal features.This method combines the advantages of denoising autoencoders and convolutional neural network,and can extract features with useful representations and robustness.In this paper,the effects of iterations,batch size,convolution kernel size,the number of convolution kernels and weighting parameters on the performance of the feature extraction method are studied.High quality signal features can be extracted through the method proposed in this paper.When the labeled data of the wireless signal data is insufficient,overfitting phenomenon occurs during the training process of supervised modulation classification,which will result in a significant decrease in the classification accuracy.The semi-supervised classification model based on autoencoders consists of two parts: unsupervised pre-training and supervised fine-tuning.It can use a large number of easily-available unlabeled data to improve network generalization performance.However,the training processes of supervised fine-tuning and unsupervised pre-training are independent from each other,the classification error and reconstruction error cannot be minimized at the same time.This will reduce the effectiveness of unlabeled data in improving network generalization performance.To solve this problem,this paper proposes a semi-supervised modulation classification model based on an improved SCDAE.In this model,unsupervised training alternates with supervised training until convergence,so that the generalization ability of unsupervised training and the classification ability of supervised training can reach the best state at the same time.The semi-supervised modulation classification model proposed in this paper improves the classification accuracy by 11% compared to the supervised model when there is insufficient labeled data and unlabeled data is sufficient.