Fast Recognition of Distributed Fiber Optic Vibration Sensing Signal based on Machine Vision in High-speed Railway Security

Abstract

Accurate and effective identification of multi-vibration events detected based on the phase-sensitive optical time-domain reflectometer (Φ-OTDR) is an effective method to achieve precise alarm. This study proposes a real-time classification method of Φ-OTDR multi-vibration events based on the combination of convolutional neural network (CNN), bi-directional long short-term memory network (Bi-LSTM) and connectionist temporal classification (CTC), which can quickly and effectively identify the type and number of vibrations contained in the data image when multiple vibration signals are present in a single image, and manual alignment is not required for model training. Noncoherent integration and pulse cancellers are used for raw signal processing to generate spatio-temporal images. CNN is used to extract spatial dimensional features in spatio-temporal images, Bi-LSTM extracts temporal dimensional correlation features, and the hybrid features are automatically aligned with the labels by CTC. A dataset of 8,000 vibration images containing 17,589 abnormal vibration events is collected for model training, validation and testing. Experiments show that the recognition model C3B3 trained with this method can achieve 210 FPS and 99.62% F1 score on the test set. The system can achieve the real-time classification of multiple vibration targets at the perimeter of high-speed railway and effectively reduce the false alarm rate of the system.