We propose a novel approach to the recognition of particular classes of non-conventional events in signals from phase-sensitive optical time-domain-reflectometry-based sensors. Our algorithmic solution has two main features: filtering aimed at the de-nosing of signals and a Gaussian mixture model to cluster them. We test the proposed algorithm using experimentally measured signals. The results show that two classes of events can be distinguished with the best-case recognition probability close to 0.9 at sufficient numbers of training samples.
We use a semiconductor laser with an erbium-doped fiber amplifier with power limited to 1 W. The wavelength of the probe signal was 1550 nm, duration of the probe pulse was 100–500 ns, coherence length of the laser is 30 km and the signal was launched into a standard single-mode telecommunication optical fiber SMF28 of length approximately 50 km. The fiber is probed on ranges of a few km. The use of multimode fibers is not needed for our purposes because it affects spatial resolution.
Since the recognition algorithm assigns to the class input events only to existed classes, it is important to exclude the effect of different types, which is beyond the classes considered here. For this, one needs to define through a preliminary test a probability threshold pc. If the algorithm predicts that an input event belongs to the first class with probability p1, this event belongs to the second type with probability p2 = 1 − p1. The workflow of the algorithm stops in cases where one of these probabilities is less than the threshold.