High-efficiency intrusion recognition by using synthesized features in optical fiber perimeter security system

Abstract

In an optical fiber perimeter security system, due to the fact that a large quantity of samples are collected in the process of data acquisition, this heavy data burden inevitably degrades the efficiency and accuracy of intrusion recognition. Hence, it is urgent to remove the redundancy of the collected data records, which essentially requires to describe event features in a concise and proper way. In this paper, we propose a synthesized feature based intrusion recognition method, which is especially suitable to describing the fiber intrusion vibration signals with both wide bandwidth and high nonstationarity. Firstly, the all-phase filter bank characterized by large sidelobe attenuation and high flexibility of coefficient configuration, is employed to parallelly divide the input signal into multiple frequency channels, from which the power values can be accurately calculated. Secondly, the crossing rate of the input signal is combined with these power values to construct a synthetical feature vector, in which both the time-domain information and frequency-domain information are incorporated together. Finally, these synthetical feature vectors are fed into a radial-basis-function network based classifier to recognize 4 common intrusions (climbing, knocking, waggling and cutting). Essentially, the high efficiency of our proposed scheme lies in the parallel pipeline mode of the configurable filter bank and simple calculation of features, which facilitates speeding up the intrusion recognition. The high accuracy of our proposed scheme lies in two aspects: 1) the all-phase filter bank possesses small inter-channel interference, which helps to reduce the inter-coupling between output power values; 2) the synthesis of both frequency-domain information and time-domain information ensures the completeness of feature description. Experiments show that the sensing range of the proposed scheme can reach 2.25 km. Moreover, compared with the empirical mode decomposition based method, the proposed method not only improves the precision, but also significantly speeds up the recognition.