Research On The Intrusion Recognition Method Based On Hybrid Feature Extraction For Optical Fiber Perimeter Security System

In recent years,due to the advantage of high sensitivity,fast response and simple,Dual Mach-Zehnder interferometry distributed sensing system has been widely used in perimeter security.However,the recognition categories of intrusions,the recognition accuracy and the performance of real-time in this system are limited by the performance of the signal processing.To realize the intrusion recognition high-efficiently and highaccuracy,three aspects should be considered: save the system cost,improve the mode of signal processing and extend the feature extraction.To save the system cost,this paper proposed an amplitude-frequency compensation based FIR filter with accurately controllable cut-off frequency.When applying this filter to the endpoint,one can find that,this filter can not only remove environment noise and thus achieve high-accuracy endpoint detection,but also save more than 70% multiplier resources and system time delay.It should be noted that,during the system idle period which occupies overwhelming majority of the sensing time,endpoint detection has to be persistently continued until an invasion happens.Hence,this method can save a lot of system cost and improve the response speed.To improve the mode of signal processing,we used the all-phase filter bank in the stage of signal pre-processing,the filter coefficients of this filter bank can be computed by some closed-form analysis expression.More importantly,compared with the existing empirical mode decomposition method of repeated iteration of the processing mode,this filter bank can work in a parallel pipeline mode and improve the efficiently.Experiments shows that,combined with RBF neural network,this method can accurately classify the two common intrusions(climbing the fence and knocking the cable).To extend perspective of the feature selection,based on the simple all-phase filter bank scheme,this paper proposed a hybrid features based intrusion recognition scheme.This scheme aims to contain comprehensive information(statistics,frequency domain and time-domain)of the intrusions signal and thus selecting the kurtosis of each channel and the zero-crossing rate of the signal to construct the hybrid feature vector.Kurtosis is suitable for non-stationary vibration signal due to its sensitivity to pulse signals,Zerocrossing rate can describe the fluctuation of the signal in a period of time.Combining these two features can deeply and comprehensively describe different intrusions.Field experiments proved that,the proposed hybrid feature extraction based scheme can accurately discriminate four common intrusions(climbing the fence,knocking the cable,waggling the fence and cutting the fence)with an averaged 94% recognition rate.Furthermore,integrating with time-domain,frequency-domain and statistics information shorten the feature vector and thus improve the efficiency.Hence,this scheme is suitable for the practical applications.