Controllable diffusion framework for imbalanced Phi OTDR events classification.

Area of Science:

• Electrical Engineering
• Data Science
• Signal Processing

Background:

• Phase-Optical Time Domain Reflectometry (Φ-OTDR) systems monitor power grids, detecting anomalies like digging or watering.
• Existing deep learning models struggle with imbalanced Φ-OTDR datasets, where abnormal events are rare compared to normal noise.
• This long-tailed distribution significantly hinders accurate classification of critical power grid events.

Purpose of the Study:

• To introduce a novel framework, Controllable Diffusion (ConDiff), for generating high-quality synthetic abnormal event data.
• To address the long-tailed imbalance problem in Φ-OTDR data classification for enhanced power grid monitoring.
• To improve the accuracy of classifying rare anomalies in real-world power grid infrastructure.

Main Methods:

• Developed the Controllable Diffusion (ConDiff) framework with three key components: Feedback-guided Φ-OTDR Augmenter, High-Quality Sample Selection, and Dynamic Threshold Adjustment.
• Utilized diffusion models to generate synthetic Φ-OTDR data simulating various abnormal events.
• Implemented a feedback loop for dynamic control and quality assessment of generated synthetic samples.

Main Results:

• The ConDiff framework demonstrated significant improvements in classification accuracy on the BJTU-OTDR-LT dataset, ranging from 3.7% to 7.2% compared to baseline methods.
• Successfully generated high-quality synthetic samples for underrepresented abnormal event classes.
• Effectively mitigated the challenges posed by the long-tailed data distribution in Φ-OTDR event classification.

Conclusions:

• The proposed ConDiff framework offers a robust solution for handling imbalanced Φ-OTDR datasets in power grid monitoring.
• Synthetic data generation using ConDiff enhances the performance of deep learning models for anomaly detection.
• This approach is crucial for improving the reliability and safety of power grid infrastructure through accurate event classification.