Event recognition method based on dual-augmentation for a Φ-OTDR system with a few training samples

Area of Science:

• Fiber Optic Sensing
• Machine Learning
• Artificial Intelligence

Background:

• Data-driven pattern recognition using neural networks is advancing intrusion event recognition in Φ-OTDR systems.
• A significant challenge is the scarcity of training samples due to difficulties in collecting intrusion signals and the time-consuming nature of data labeling.

Purpose of the Study:

• To address the limited training data problem in Φ-OTDR intrusion event recognition.
• To propose a few-shot learning classification method that enhances sample availability for neural network training.

Main Methods:

• A few-shot learning classification approach is presented, incorporating time series transfer.
• Cycle Generative Adversarial Network (CycleGAN) is utilized for data augmentation to expand rare intrusion event samples.
• The augmented dataset is used to meet the requirements for neural network training.

Main Results:

• The proposed method effectively expands the dataset, enabling network training even with minimal samples.
• Achieved an average accuracy of 90.84% on the validation set across 5 classification tasks.
• Demonstrated a classification accuracy of 79.28% for minor classes, even when only two samples were available for training.

Conclusions:

• The few-shot learning method combined with CycleGAN data augmentation successfully overcomes the challenge of limited training data in Φ-OTDR systems.
• This approach significantly improves intrusion event recognition accuracy, particularly for rare events.
• The method offers a viable solution for enhancing the performance of fiber optic sensing systems in real-world applications.