Unified self-training framework for open-set semi-supervised event recognition in distributed optical fiber sensing.

Area of Science:

• Optical Physics
• Sensor Technology
• Machine Learning

Background:

• Distributed optical fiber sensing (DOFS) offers extensive monitoring capabilities but faces challenges with data requirements and recognizing novel events.
• Current event recognition models struggle in open environments due to reliance on large labeled datasets and limitations in identifying unseen disturbances.

Purpose of the Study:

• To develop an open-set semi-supervised event recognition method for DOFS systems.
• To address the limitations of existing models in handling scarce labeled data and unknown events in real-world applications.

Main Methods:

• A unified self-training framework integrating closed-set and open-set learning.
• A buffer-based strategy for selecting reliable unknown samples.
• Simultaneous optimization for recognizing both known and unseen events.

Main Results:

• Achieved 95.14% open-set accuracy in event recognition.
• Demonstrated superior performance compared to baseline methods like IOMatch and OpenMatch under low-supervision (1% labeled data).
• Showcased strong and consistent performance in both closed-set and open-set scenarios.

Conclusions:

• The proposed method significantly enhances the recognition performance and robustness of DOFS systems.
• Provides a reliable foundation for intelligent optical fiber sensing in dynamic, open environments.
• Overcomes limitations of data scarcity and unknown event detection in practical DOFS applications.