Extraction Signal Fingerprint Based On Deep Learning Technology

The signal fingerprint aiming at capturing the differences carried by transmitted wireless signal caused by the non-ideal component in transmitting circuit or the structure of transmitting circuit. The traditional signal fingerprint extraction method is based on pre-designed formula.This results in high requirement for priori information and narrow application range. The solution to previous problems is provided in this thesis. The deep learning technology is used for extracting fingerprint of signal based the given wireless communication signal. The research focus on the signal fingerprint extraction based deep learning network, deep learning network optimization and wireless devices recognition. The major contributions of this thesis are as follows:1. The signal fingerprint extraction methods based on deep learning network is poroposed. Actually, the convolutional deep belief network(CDBN) is utilized for extracting fingerprint of signal. The CDBN model,signal pre-treatment process and the signal fingerprint extraction based CDBN is introduced. The simulation results suggest that this method is feasible and the extracted signal fingerprint provides good performance.2. The CDBN utilized for extracting fingerprint is optimized to reduce the fingerprint extraction time and enhance its ability for distinguishing wireless devices. The structure and the training objective of convolutional restricted Boltzmann machine (CRBM) which is the element of CDBN is optimized at first. This leads to the improvement of extracted signal fingerprint. After that the parameter of CDBN and CRBM is optimized in order to decrease the time consumed for signal fingerprint extraction. During the parameter optimization process, the existing evaluating indicator or designed evaluating indicator is used for selecting recommendation parameter. In the end, the proposed setting value of each parameter is exhibited.3. The feasibility and the resolution ability of extracted signal fingerprint are evaluated under several simulation scenarios and real world application scenario based on the accuracy of wireless devices recognition. Three simulation scenarios are proposed aiming at verifying the feasibility of signal fingerprint based on CDBN and providing the evaluating indicator for parameter choosing, evaluating the feasibility and the resolution ability of extracted signal fingerprint under approximate real-world application scenario, verifying the feasibility of signal fingerprint based on CDBN when the traditional signal fingerprint extraction is invalid. Analysis the wireless devices recognition with given sample signal set and without given sample signal set. The results suggests that under different scenario, when signal noise ratio (SNR) is relative low, the resolution if extracted signal fingerprint based on CDBN enjoys good resolution ability and the accuracy of wireless devices is above 70%. In real-world application, signal fingerprint extracted based on CDBN performs well and can guarantee high wireless devices recognition accuracy. And when traditional signal fingerprint extraction is invalid the signal fingerprint extraction based on CDBN is valid. Besides,the physical significance of extracted signal fingerprint is briefly analyzed utilizing the real-world wireless communication signal.