Wireless Communication Signal Sample Labeling Technology For Individual Identification Applications

With the increasing development of wireless communication technology,the modern electromagnetic environment is becoming more and more complex,the identification technology of radio frequency fingerprinting has replaced the traditional modulation identification technology as the main method of solving the individual identification of radiation sources at present.Focus of the radiation source individual identification technique is to find the fingerprint feature that uniquely represents the radiation source individual.The differences generated in the physical layer of wireless communication transmitter devices are usually used as fingerprint features of the device to distinguish between different devices,but due to the increasing level of electronic component manufacturing processes,the differences between components under the same batch are gradually becoming smaller and fingerprint feature extraction becomes difficult.Some experts and scholars have researched and proposed the use of deep learning technology to solve the problem of individual identification of RF devices,so deep learning technology has been widely used and developed in the area of individual identification of radiation sources.Deep learning networks for individual radiation source identification require a large number of labeled samples to be trained,however,in practice,there are large sets of training samples that lack annotation and in some cases only a small number of labeled samples exist in the individual radiation source identification process.How to find an effective way to fully apply the training sample set,annotate the sample set with individuals,and improve the accuracy of the radiation source individual recognition results is now a class of problem that needs to be solved.In this paper,unsupervised learning-based and semi-supervised with contrastive learning-based methods are designed for the annotation of communication radiation source samples under the condition of unlabeled samples and a small number of labelled samples,respectively.The main research work of this paper is as follows.1.Starting from the mechanism of fingerprint formation in wireless communication signal transmitters,introducing the general design structure of transmitters and studying the influence of signal fingerprints brought by the main hardware of I/Q modulators,IF crystal oscillators and power amplifiers respectively,giving the influence factors and types of features of fingerprint characteristics;2.Fingerprint features are extracted from the carrier frequency deviation of the steady-state signals of wireless communication signals.By analyzing the fingerprint feature extraction method based on short-time Fourier variation,the fingerprint feature extraction method based on Hilbert yellow transform,and the fingerprint feature extraction method based on higher-order accumulation,the carrier frequency components of the wireless communication signal samples are finally decomposed by the empirical modal algorithm in combination with the experimental results;3.Based on the steady-state signal carrier frequency deviation feature extraction method,the paper proposed an unsupervised sample annotation model,including feature compression of the feature vector and individual recognition.The proposed feature compression module reduces the impact of high-dimensional feature vectors on the subsequent recognition work,reduces the complexity of the individual recognition classification process and improves the accuracy of the individual recognition classification results to a certain extent;4.A semi-supervised sample annotation model is proposed to improve the performance of individual recognition classifiers using feature enhancement for labelled signals with a small number of labelled samples.The proposed feature enhancement module effectively extracts the"difference" features from the feature vectors through the idea of "desimilarity",and "purifies" the features to reduce the classification difficulty for subsequent individual recognition.