Ensemble Learning for Detection and Classification of Noisy Nonstationary Signals

In this work, the signal recognition problem is considered on the data obtained from phase-sensitive optical time-domain reflectometer developed in Scientific Educational Center Photonics and IR-Technology, Bauman Moscow State Technical University. The signals are divided into three classes of events: steps, digging and absence of event, which includes noises and background signal. The event recognition problem consists of the following steps: 1) determination of the presence of an event by the type of signal (localization), 2) construction of training set and 3) classification. For solving the problem, fourier transformation, convolutional neural networks (CNN) and gradient tree boosting (XGBoost) are applied. Classification model based on the stacking of CNN and XGBoost models was developed. Recognition accuracy of the model achieves 97.2%.