Pattern recognition in distributed fiber-optic acoustic sensor using an intensity and phase stacked convolutional...

Area of Science:

• Geophysics
• Sensor Technology
• Machine Learning

Background:

• Distributed acoustic sensors (DAS) detect vibrations with high spatial resolution.
• Advanced algorithms are crucial for classifying vibration events in DAS data.
• Deep convolutional neural networks (CNNs) are suitable for DAS event classification but require large datasets.

Purpose of the Study:

• To improve the classification accuracy of vibration events in DAS data.
• To address the challenge of limited training data for data-driven DAS methods.
• To enhance the utilization of collected information for CNNs in DAS applications.

Main Methods:

• Proposed an intensity and phase stacked CNN (IP-CNN) to leverage both intensity and phase information from DAS.
• Employed data augmentation techniques to enlarge the training dataset size.
• Investigated the impact of various data augmentation methods on CNN performance.

Main Results:

• The proposed IP-CNN with data augmentation achieved 88.2% classification accuracy on a 1km DAS dataset.
• Utilizing both intensity and phase information improved classification performance.
• Data augmentation effectively increased the training dataset size and model accuracy.

Conclusions:

• Combining intensity and phase information with data augmentation significantly enhances DAS classification accuracy.
• The proposed IP-CNN method offers a powerful approach for DAS pattern recognition in real-world applications.
• Enlarging training datasets through augmentation is vital for maximizing CNN performance in DAS.