Distributed Optical Fiber Sensing Event Recognition Based on Markov Transition Field and Knowledge Distillation

Abstract

We propose a distributed optical fiber sensing event recognition scheme based on Markov Transition Field (MTF) and knowledge distillation. The event recognition algorithm has the advantages of being lightweight, fast, and high accuracy. The event data are converted into images by the MTF algorithm, which keeps the event signals correlated in the time domain while highlighting the visual differences between different events. A two-stage knowledge distillation model compression method is proposed, which effectively compresses the large-scale model into a lightweight model with optimal learning capability, ensuring the lightweight and efficient recognition of events by the compressed model (student model). The experimental results show that the student model improves the recognition rate of six events by 5.2% and achieves 96.6% event recognition accuracy by the two-stage knowledge distillation method. The size of the student model is only 1.4 MB, the number of parameters is only 0.35 M, and the FLOPs are only 0.17 G. The student model recognizes a single event in 0.129s on a low configuration device, which can meet the requirements of deployment and real-time monitoring of resource-limited devices.

Keywords

• Distributed optical fiber sensing
• Markov transition field
• knowledge distillation
• sensor recognition
• deep learning