Phase sensitive optical time domain reflectometer (φ-OTDR) can retrieve vibration waveforms based on linear relationship between phase change and external events. Yet, it is difficult to identify different events due to the complexity of working environment. How to accurately determine the type of vibration events and thus reduce false alarm rate is important in many practical engineering applications. The existing deep learning (DL) algorithm can directly extract the original data feature, without manual extraction. Hence, DL is usually used to classify and recognize multiple events in φ-OTDR. In this work, a dual input deep convolutional neural network (Di-DCNN) is applied to distinguish six kinds of actual vibration events (including walking, tapping, blowing and raining, vehicle passing, digging and background noise). The features of these two inputs are extracted, respectively, and fused together to identify six vibration events. For comparison, network models with other five inputs are employed for event recognition, including single input of 1D time-demain or 2D image of phase (amplitude) data, and dual input of 1D time-demain and 2D image of amplitude data. Here, the 2D image denotes the transformation of 1D data by Markov Transition Field (MTF). Experimental results show that the Di-DCNN with 1D time-domain phase waveforms and its 2D MTF image being the two inputs considerably improve the recognition accuracy. The average recognition rate of six kinds of vibration events is higher than 94%.
|