Distributed acoustic sensors (DASs) have the capability of registering faint vibrations with high spatial resolution along the sensing fiber. Advanced algorithms are important for DAS in many applications since they can help extract and classify the unique signatures of different types of vibration events. Deep convolutional neural networks (CNNs), which have powerful spectro-temporal feature learning capability, are well suited for event classification in DAS. Generally, these data-driven methods are highly dependent on the availability of large quantities of training data for learning a mapping from input to output. In this work, to fully utilize the collected information and maximize the power of CNNs, we propose a method to enlarge the useful dataset for CNNs from two aspects. First, we propose an intensity and phase stacked CNN (IP-CNN) to utilize both the intensity and phase information from a DAS with coherent detection. Second, we propose to use data augmentation to further increase the training dataset size. The influence of different data augmentation methods on the performance of the proposed CNN architecture is thoroughly investigated. The experimental results show that the proposed IP-CNN with data augmentation produces a classification accuracy of 88.2% on our DAS dataset with 1km sensing length. This indicates that the usage of both intensity and phase information together with the enlarged training dataset after data augmentation can greatly improve the classification accuracy, which is useful for DAS pattern recognition in real applications.
ASJC Scopus subject areas
- Atomic and Molecular Physics, and Optics