Markov Transition Fields and Deep Learning-Based Event-Classification and Vibration-Frequency Measurement for φ-OTDR

In this paper, a novel method, relying on Markov Transition Fields (MTF) and deep learning, is proposed to classify the vibration-events and measure vibration-frequency, for phi-OTDR based fiber-optic distributed vibration sensor. The normalized time series of a signal detected by phi-OTDR are converted into an image of MTF, serving as a sample for supervised learning. Next, the MTF image is classified using a convolutional neural network (CNN) and a fully connected neural network. Initially, five different vibration-events including blowing, rain, direct and indirect knocking, and false vibration caused by noises, are classified. Furthermore, approximate single-frequency vibrations (with different central-frequency) are regarded as different classifications, where vibration frequency can be obtained by classification tasks. Compared to conventional method based on phase-demodulation with complex techniques and high cost, the proposed method demonstrates a low-cost and fast-time method. Moreover, learning algorithm is trained and tested through the data sets generated by experiments. To evaluate feasibility and validate the performance of classification, we analyze the test accuracy of classification and relevant receiver operating characteristic curves. The results indicate that our method is effective for the classification of both vibration-events and single-frequency vibrations. We believe that this work provides not only a technical reference for the application of deep learning in measuring the vibration frequency of phi-OTDR using software-technique, but also an impressive inspiration for event-classification of vibrations by a combination of image-processing methods.

Automated summary

This paper proposes a novel method using Markov Transition Fields and deep learning to classify vibration events and measure vibration frequency. The method converts the normalized time series of a signal detected by phi-OTDR into MTF images, which are then classified using convolutional neural networks and fully connected neural networks. The experimental results demonstrate the effectiveness of this low-cost and fast method.