Deep Learning Based Multi-threat Classification for Phase-OTDR Fiber Optic Distributed Acoustic Sensing Applications

This paper presents a distributed acoustic sensing based linear asset protection system along with novel signal processing and threat classification techniques. The sensing system is realized by direct detection phase-OTDR (optical time domain reflectometry). An effective signal preprocessing approach for noise reduction that aims to improve the threat detection capability of the system is proposed. The proposed method is not limited to direct detection based systems and is applicable to any phase-OTDR system. A novel deep learning based threat classification method is presented to identify various types of threats. The method uses a deep convolutional neural network trained with real sensor data. Experiments are conducted with an ITU-T G.652 fiber optic cable buried at one meter depth. The effects of applied preprocessing methods on both threat detection and threat classification performance are analyzed. The proposed preprocessing method is compared with the methods commonly used in the literature such as time differencing and wavelet denoising. The results show that by applying the proposed signal conditioning, event detection and classification methods, threat classification accuracies above 93% can be achieved with six typically observed activities, namely, walking, digging with pickaxe, digging with shovel, digging with harrow, strong wind and facility noise caused by water pipes, generators or air conditioning, at ranges of up to 40 km. The proposed classification strategy can easily be generalized for identifying different types of threats that are of interest in various security applications.