Apart and A Part: Overlapped vibration recognition for distributed optical fiber sensing based on deep learning separation
It has been proven feasible to utilize phase-measuring phase-sensitive optical time-domain reflectometry (Φ-OTDR) based acquisition instruments for collecting and classifying vibrations on distributed fibers, based on existing research. However, existing methods do not perform well in detecting multiple vibrations that occur simultaneously and estimating their categorization. In this article, we introduce a neural network approach based on Conv-TasNet to separate vibration signals. The experiment shows that the separated signals from the overlapped events can work the same as the real detected single event data in the classification test, with only a 2.4 % classification accuracy gap. The separated overlapping signals can be correctly classified into specific event types with acceptable classification accuracies.