Unsupervised Anomaly Detection Applied to Φ-OTDR.

Area of Science:

• Fiber Optic Sensing
• Signal Processing
• Machine Learning

Background:

• Distributed acoustic sensors (DAS) based on Φ-OTDR detect mechanical events via light-matter interactions in optical fibers.
• High sensitivity leads to reduced signal-to-noise ratio, necessitating advanced processing techniques.
• Current methods struggle with noise, hindering accurate event detection.

Purpose of the Study:

• To propose and evaluate an unsupervised anomaly detection method for noise reduction in Φ-OTDR signals.
• To leverage deep learning concepts for enhanced signal processing in DAS.
• To demonstrate the efficacy of the method using real-world data.

Main Methods:

• Implementation of an unsupervised anomaly detection algorithm based on deep learning.
• Training the model exclusively on event-free Φ-OTDR data.
• Testing the method's performance on real-world distributed acoustic sensing data.

Main Results:

• Significant noise reduction observed in Φ-OTDR signals after applying the unsupervised anomaly detection method.
• Improved isolation of mechanical events within the processed signals.
• Promising performance demonstrated with real-world data, validating the approach.

Conclusions:

• Unsupervised deep learning anomaly detection offers an effective solution for noise reduction in Φ-OTDR signals.
• The proposed method enhances the reliability of distributed acoustic sensing systems.
• No human-labeled data is required, simplifying the training process.