OTDR data augmentation and event detection using GAN under imbalanced samples

With the expansion of fiber optic networks, the demand for fiber health monitoring and maintenance has become increasingly prominent. As a core device for optical fiber link detection, the Optical Time Domain Reflectometer (OTDR) plays a crucial role in fault localization and loss measurement. However, OTDR faces challenges of scarce real anomaly samples and class imbalance in practical applications, leading to high false positive rates in detection models and reduced monitoring efficiency. To address these issues, this paper proposes an OTDR event detection method using a dynamic dropout-based generative adversarial network (DDW-GAN) for data augmentation. First, OTDR data is segmented and subjected to wavelet packet-based time-frequency feature extraction. Subsequently, the adversarial learning mechanism of DDW-GAN is employed to generate high-quality reflective event and non-reflective event samples, resolving the data imbalance problem. Finally, comparative experiments involving multiple classifiers—including Support Vector Machine (SVM), Random Forest, and Deep Neural Network (DNN)—are conducted. Experimental results demonstrate that the proposed method effectively mitigates the imbalance in OTDR data and achieves reliable anomaly event detection under imbalanced data conditions.