Pattern recognition using self-reference feature extraction for φ-OTDR.

Area of Science:

• Fiber optic sensing
• Signal processing
• Machine learning

Background:

• Phase-sensitive optical time-domain reflectometry (φ-OTDR) is crucial for distributed sensing.
• Accurate vibration pattern recognition is essential for φ-OTDR applications.
• Existing methods may struggle with complex vibration signatures and noisy data.

Purpose of the Study:

• To develop an advanced pattern recognition method for φ-OTDR using self-reference features.
• To enhance vibration classification accuracy for diverse events and industrial machinery.
• To identify optimal machine learning algorithms and feature extraction techniques for φ-OTDR data.

Main Methods:

• Utilized self-reference features derived from light amplitude-time-space sequences in φ-OTDR.
• Employed machine learning algorithms (kNN, Decision Tree, Random Forest, LightGBM, CatBoost) for classification.
• Optimized sample length (10 data points) and selected eight key eigenvalues for feature extraction.
• Developed a filtering algorithm to handle abnormal signals and improve robustness.

Main Results:

• Achieved an average recognition rate of 98% for common events (tapping, bending, trampling, blowing).
• Demonstrated high accuracy (93.5% average) in recognizing vibrations from five types of mining equipment.
• Identified CatBoost as the optimal machine learning algorithm for this application.
• Validated the effectiveness of self-reference features and the filtering algorithm in improving accuracy.

Conclusions:

• The proposed self-reference feature-based pattern recognition method significantly enhances φ-OTDR accuracy.
• The CatBoost algorithm combined with optimized features provides a robust solution for vibration analysis.
• This method shows strong potential for real-world applications in event detection and industrial equipment monitoring.