Label-Free Anomaly Detection Using Distributed Optical Fiber Acoustic Sensing.

Area of Science:

• Signal Processing
• Machine Learning
• Sensor Technology

Background:

• Anomaly detection in distributed optical fiber acoustic sensing (DAS) is crucial but challenging due to data scarcity and imbalance.
• Supervised learning methods are limited by the inability to catalog all potential anomalies.

Purpose of the Study:

• To develop an unsupervised deep learning method for anomaly detection in DAS that learns solely from normal data.
• To address the limitations of supervised learning in handling imbalanced and incomplete anomaly datasets.

Main Methods:

• Utilized a convolutional autoencoder to extract features from DAS signals.
• Employed a clustering algorithm to identify normal data feature centers.
• Developed a distance-based method to classify new signals as anomalous or normal.

Main Results:

• Achieved a threat detection rate of 91.5% in a high-speed rail intrusion scenario.
• Demonstrated a 5.9% higher detection rate compared to state-of-the-art supervised networks.
• Reduced the false alarm rate to 7.2%, which is 0.8% lower than supervised methods.
• Significantly decreased model parameters to 1.34 K using a shallow autoencoder.

Conclusions:

• The proposed unsupervised deep learning method offers a robust and efficient solution for anomaly detection in DAS.
• The approach outperforms supervised methods in terms of detection rate and false alarm rate while being computationally less intensive.
• This method provides a viable alternative for real-world applications where anomaly data is limited.