Pattern Recognition of Phase-Sensitive Optical Time-Domain Reflectometer Based on Conditional Generative Adversarial Network Data Augmentation

Objective We aim to address limited data acquisition in fiber optic sensing technology, especially in phase-sensitive optical time-domain reflectometry. A data augmentation method based on conditional generative adversarial networks (GANs) is proposed to generate a large number of training samples and improve the detection capability and performance of the classifier model. Methods The experimental data collection is conducted using a phase-sensitive optical time-domain reflectometer (Phi-OTDR). First, the collected real data are adopted as input to the conditional GAN. The GAN model automatically extracts signal features and generates realistic signal data with the assistance of input conditions, with the specific experimental flow shown in Fig. 7. Second, the generated data and original data are separately fed into classifiers such as decision trees, support vector machines, and convolutional neural networks for classification. By comparing the detection results of the generated and raw data across different classifiers, the effectiveness of the data augmentation method is evaluated, and the specific comparison results are shown in Fig. 12. This comprehensive approach can assess the influence of the generated data on the classifier performance to address limited data acquisition in fiber optic sensing technology. Results and Discussions The experimental results demonstrate that the detection results of the generated data significantly improve across decision trees, support vector machines, and convolutional neural networks. The generated data enhance the detection capability and performance of the classifier models, achieving the target identification in Phi-OTDR. Furthermore, improvements in the conditional GAN can generate more realistic signal data, further enhancing the model performance. Conclusions We successfully address the data acquisition limitations in Phi-OTDR by a data augmentation method based on conditional GAN. The generated data improve the detection capability and performance of the classifier models. The research findings provide new insights and methods for small-sample detection, and also valuable references for the applications of other fiber optic sensing technologies.