Abstract
The application of the Φ-OTDR (Phase-Optical Time Domain Reflectometry) system in real-time monitoring of power grid infrastructure has been proven effective in identifying and classifying various anomalies, such as digging, watering, and shaking. However, previous deep learning-based methods for Φ-OTDR event classification are primarily designed for balanced classification problems, where the number of abnormal and normal event samples is relatively equal. In practical scenarios, the data for abnormal events are often much smaller than those for normal events (noise), resulting in a long-tailed distribution problem that poses significant challenges for accurate classification. To address this long-tailed imbalance issue in the practical application of Φ-OTDR data, we introduce the Controllable Diffusion (ConDiff) framework, which aims to generate high-quality synthetic samples for abnormal situations. The ConDiff framework is composed of three essential components: Feedback-guided Φ-OTDR Augmenter, the High-Quality Sample Selection module, and the Dynamic Threshold Adjustment module. The Feedback-guided Φ-OTDR Augmenter utilizes diffusion model to generate synthetic samples that simulate abnormal events. The High-Quality Sample Selection module evaluates the quality of the generated synthetic samples and selects high-Quality samples. The Dynamic Threshold Adjustment module provides real-time feedback to dynamically control the sample generation process of Feedback-guided Φ-OTDR Augmenter. Compared to current state-of-the-art baselines, our proposed ConDiff framework achieves a notable improvement in classification accuracy, with an increase ranging from 3.7% to 7.2% in the BJTU-OTDR-LT dataset. This improvement demonstrates the effectiveness of the proposed ConDiff framework in addressing the long-tailed imbalance problem in Φ-OTDR event classification. The code will be released upon acceptance.