Research On Deep Learning Based Open-set Recognition Technology For Wireless Signals

With the rapid development of wireless technology and smart home applications,the communications on the unauthorized frequency ISM(Industrial Scientific Medical)band has become extremely crowded,and the spectrum resources have become increasingly tight,which presents severe challenges for the efficient sharing of spectrum resources,interference diagnosis and management of communication networks,etc.Wireless Signal Recognition(WSR)is a general term for the recognition of wireless technology protocols or equipment types through the analysis and processing of wireless air interface signals.It is also an important technical means for cognitive wireless communication,wireless spectrum monitoring and interference management.The classical WSR technology usually uses the method of pattern recognition to extract artificial features based on the close-set hypothesis,and trains a classifier to do the inference of signal type.However,the heterogeneous diversity of wireless technology and the open dynamic nature of wireless communication environment usually lead to the failure of the close-set hypothesis in the real world,and the open-set application environment presents severe technical challenges to the performance and difficulty of WSR.In this thesis,the problem of open-set recognition(OSR)for wireless signals is transformed into an N+l classification problem(where N represents the number of known signal types and 1 represents the other unknown signal types),and the WSR method based on data-driven end-to-end deep learning in ISM band is studied to improve the performance of open-set recognition.Specifically,the research work of this paper can be divided into the following three aspects:1.Based on the intra-class splitting(ICS)algorithm,this thesis deeply studied the construction method of decision boundary samples,and proposed an improved OSR algorithm for wireless signals based on decision boundary samples.Firstly,the algorithm designs a classifier for N classes of known signals,constructs new decision boundary samples from its decision domain to mimic the unknown class,and finally designs and trains an N+1 classifier to realize the OSR of wireless signals.Among them,the construction of decision boundary samples mainly includes two parts:one is the splitted samples with high entropy score that are separated from the known N classes with the aid of ICS idea;In the other part,an iterative method is designed to generate adversarial samples near the decision boundary to enhance the decision domain of unknown class by means of adversarial learning.In order to improve the recognition accuracy,in the training stage,this paper simultaneously trains the N+1 open-set classifier and the N class close-set classifier,while only the former is used in the inference stage.Finally,the proposed algorithm is tested on the dataset with ten kinds of wireless signals collected in ISM band,including Wi-Fi,Bluetooth,ZigBee and microwave oven,as well as three public image datasets,and the resulting effectiveness of the proposed algorithm is analyzed and verified accordingly.2.Combined with multi-task learning(MTL)and based on ICS algorithm,this thesis further proposes a wireless signal OSR algorithm based on multi-task learning to improve the robustness of OSR process.An MTL network is built to train wireless signal recognition and modulation recognition simultaneously through a hard parameter sharing pattern.In the process of decision domain learning,in order to reduce the influence of splitting ratio hyperparameter on algorithm performance,an adaptive threshold adjustment strategy is adopted to divide the original data into inner samples and outer samples according to confidence degree,and the outer samples are collected to simulate the unknown class.In order to solve the problem of unbalanced amount of outer samples,this thesis also uses data augmentation algorithm for outer samples to improve training performance.Finally,the performance test and experimental analysis of the proposed algorithm are carried out on one wireless signal dataset collected in ISM band and one public modulation recognition dataset to verify its effectiveness.3.Based on PyQt,an online wireless signal recognition software with OSR algorithm embedded as the key module is developed and implemented on personal computer(PC)platform.By analyzing the functional requirements of the WSR software,this thesis elaborately demonstrates the related functional modules in details based on the multi-threading mechanism,and carefully tests the basic functions and real-time performance of the software.The experimental results reveal that the average detection delay of a 4096-point sample is about 120ms.Finally,the two proposed OSR algorithms are individually tested online in four indoor scenarios to verify their effectiveness.