Real-Time Multi-Class Disturbance Detection for Φ-OTDR Based on YOLO Algorithm

Weijie Xu, Feihong Yu, Shuaiqi Liu, Dongrui Xiao, Jie Hu, Fang Zhao, Weihao Lin, Guoqing Wang, Xingliang Shen, Weizhi Wang, Feng Wang, Huanhuan Liu, Perry Ping Shum, Li-Yang Shao2022 · DOI

TL;DR

This paper introduces a novel real-time, multi-class disturbance detection system for fiber optic sensing (Φ-OTDR) by adapting the YOLO object detection algorithm. By treating vibration data as images, the system achieves high accuracy in identifying and classifying various intrusion events, significantly outperforming traditional methods in speed and enabling practical, continuous monitoring.

What Problem Does This Paper Address?

Distributed Optical Fiber Sensing (DOFS) systems, particularly those using Phase-Sensitive Optical Time-Domain Reflectometry (Φ-OTDR), are powerful tools for long-distance vibration monitoring. They can pinpoint the location and temporal characteristics of disturbances. However, accurately and quickly classifying different types of disturbances in real-time has been a significant challenge. Traditional methods often rely on human-engineered features, which are brittle and fail in complex, real-world scenarios. While deep learning approaches like CNNs offer better robustness, they can be computationally intensive, hindering real-time performance. Existing object detection methods, even those adapted for this domain, have often been limited to single-class detection or lack the speed required for continuous monitoring. This paper aims to bridge this gap by developing a fast, accurate, and multi-class disturbance detection system for Φ-OTDR.

Methods

The core of this work is the adaptation of the You Only Look Once (YOLO) object detection algorithm for Φ-OTDR data. The authors propose a workflow that transforms the raw, one-dimensional sensing data into two-dimensional "spatial-temporal sensing images." This transformation allows them to leverage YOLO's capabilities.

Here's a breakdown of their approach:

  • Data Transformation: Raw Rayleigh backscattering (RBS) traces from the Φ-OTDR system are collected and arranged sequentially to form a 2D matrix. This matrix, representing the spatial and temporal evolution of vibrations, is then normalized and converted into an image where pixel intensity corresponds to amplitude.
  • YOLO Architecture: The paper utilizes a YOLO-based network, specifically employing the Darknet53 backbone for feature extraction and Feature Pyramid Networks (FPN) for multi-scale feature fusion. This architecture is known for its efficiency in performing localization and classification in a single stage.
  • One-Stage Detection: Unlike traditional two-stage detectors (e.g., R-CNN, Fast R-CNN, Faster R-CNN) that first propose regions and then classify them, YOLO performs both localization (bounding box prediction) and classification simultaneously within a single network. This significantly speeds up processing.
  • Transfer Learning: To enhance performance and reduce training time, the YOLO network is pre-trained on the ImageNet dataset. The initial layers' weights are retained, and the rest are randomly initialized for the Φ-OTDR dataset.
  • Multi-Class Classification: The system is designed to distinguish between five specific types of disturbance events: calm state, rigid collision, hitting the protective net, shaking the protective net, and cutting the protective net.

Datasets

The authors collected a new dataset specifically for this research:

  • Custom Dataset: Five types of disturbance events were simulated and recorded using a 1.6 km sensing fiber. The dataset comprises a total of 5787 samples.
  • Calm State (I): 800 samples
    • Rigid Collision (II): 743 samples
    • Hit Net (III): 1932 samples
    • Shake Net (IV): 1707 samples
    • Cut Net (V): 605 samples
  • ImageNet: Used for pre-training the YOLO network, a standard practice in deep learning for image-based tasks.

Key Insights and Ideas

  • Treating Vibration Data as Images: The most significant insight is the successful application of object detection techniques, designed for visual images, to spatio-temporal vibration data from Φ-OTDR. This opens up a new paradigm for analyzing such data.
  • YOLO's Speed Advantage: The one-stage nature of YOLO is crucial for achieving real-time performance, a critical requirement for continuous monitoring systems. The paper quantifies this speedup against established two-stage methods.
  • Decoupling Location and Type: The authors emphasize the importance of separating disturbance type from its location to avoid misclassifications due to inherent spatial patterns. Their approach aims to achieve this.
  • "Calm State" Handling: A specific strategy is employed to treat the "calm state" as a baseline, ensuring that only actual disturbance events are highlighted, reducing visual clutter and focusing on anomalies.

Key Results

  • High Accuracy: The YOLO-based scheme achieved an accuracy of 96.14% for multi-class disturbance detection.
  • Real-Time Performance: The system demonstrated a processing speed of 22.83 FPS, which is significantly faster than:
  • 44.90 times faster than Fast-RCNN.
  • 3.79 times faster than Faster-RCNN.
  • Processing Time: For a 0.5-second sensing data segment (treated as an image), the processing time was approximately 0.0438 seconds. This is less than the sliding step used for video generation (0.05 seconds), enabling real-time video processing.
  • Multi-Event Detection: The system successfully detected multiple events occurring simultaneously in the sensing video.
  • Robustness: The method showed good performance in correctly locating events, with a low false alarm rate.

Limitations and Future Work

  • Dataset Size: While the dataset is substantial for this specific application, the authors acknowledge that larger and more diverse datasets could further improve robustness.
  • "Calm State" Misclassification: The paper notes that while localization is generally accurate, the "calm state" can sometimes be misclassified as an event if not handled carefully, though their specific treatment mitigates this.
  • Future Work: The authors suggest that their method is well-suited for online monitoring scenarios and could be further optimized for even higher speeds or more complex disturbance types. The potential for integrating this into broader security or industrial monitoring systems is implied.

Summary updated

BibTeX
@misc{xu2022real,
  author = {Weijie Xu and Feihong Yu and Shuaiqi Liu and Dongrui Xiao and Jie Hu and Fang Zhao and Weihao Lin and Guoqing Wang and Xingliang Shen and Weizhi Wang and Feng Wang and Huanhuan Liu and Perry Ping Shum and Li-Yang Shao},
  title = {{Real-Time Multi-Class Disturbance Detection for Φ-OTDR Based on YOLO Algorithm}},
  year = {2022},
  doi = {10.3390/s22051994},
}
Figure