A Data Augmentation Method Based on Time Domain Convolutional Mask Network for improved event recognition in φ-OTDR

In the φ-OTDR sensing system, addressing limited training samples presents a significant challenge, but it is not the only issue we encounter. In real-world application scenarios, the data collection environment is often affected by interference from multiple sources of events. This type of event noise can severely impact the performance of the recognition system. To tackle this problem, we propose a data augmentation method based on a time-domain convolutional mask network. This approach not only facilitates data augmentation but also introduces a unique type of event noise that enhances the performance of the recognition system, an aspect that previous data augmentation methods have not addressed. Experimental results indicate that in the classification task, our proposed method improves classification accuracy from 84.7% to 88.6%, highlighting the potential of deep learning techniques in processing complex vibration signals.