Effective zero-shot learning method for event classification in Φ-OTDR sensing systems

Area of Science:

• Fiber optic sensing
• Machine learning for security
• Signal processing

Background:

• Traditional Φ-OTDR intrusion detection requires extensive training data, leading to overfitting on common events.
• Difficulty in simulating or acquiring sufficient samples for rare intrusion events poses a significant challenge.
• Existing methods struggle with generalization when faced with novel or infrequent intrusion types.

Purpose of the Study:

• To develop a zero-sample learning model for recognizing one-dimensional intrusion event signals with insufficient training data.
• To address the overfitting issue prevalent in conventional intrusion detection systems.
• To enhance the robustness and applicability of Φ-OTDR systems for diverse intrusion scenarios.

Main Methods:

• Proposed a zero-sample learning one-dimensional residual model (APL-ZSL-1DResNet).
• Incorporated attribute point loss (APL) to improve feature representation for unseen classes.
• Validated the model on both a self-made and an open dataset, treating each category as zero-sample events.

Main Results:

• Achieved an average recall rate of 75% and 66% for zero-sample intrusion events across two datasets.
• Maintained high average recall rates of 94.6% and 83.5% for common intrusion events.
• Demonstrated significant improvement in recognizing intrusion events with limited or no prior training samples.

Conclusions:

• The APL-ZSL-1DResNet model effectively handles scenarios with insufficient training data for intrusion event recognition.
• Zero-sample learning offers a viable solution to overcome the limitations of data-hungry models in Φ-OTDR systems.
• The proposed method enhances the practical deployment of fiber optic sensing for security applications by improving detection of rare events.