Real-Time Classification of Anthropogenic Seismic Sources from Distributed Acoustic Sensing Data: Application for Pipeline Monitoring

Abstract Distributed Acoustic Sensing (DAS) is an innovative method to record acoustic waves using an optical fiber as a network of sensors. Current DAS devices can monitor up to 50 km of optical fiber and the use of optical repeaters can raise even more this length, while allowing a spatial discretization of the order of a meter. Handling such amount of data is a challenge in terms of data management and data analysis (such as event source identification), more specifically for monitoring applications such as infrastructures or natural hazards. In this work, we propose a processing chain for real-time classification of anthropogenic sources using a combination of Random Forest (RF) and Random Markov Field (RMF). To develop the method, we choose to focus on the application of pipeline monitoring. The algorithm is therefore trained to recognize six classes of seismic sources: pedestrian, impact, backhoe, compactor, leak, and noise. All the sources were triggered and recorded on our own test bench under controlled conditions. The average sensitivity of our processing chain reaches 83% with the use of only RF and achieves 87% in combination with RMF. Classification maps show that the RMF approach can increase the average sensitivity by removing isolated signals. In addition to this improvement in sensitivity, this new approach also permits to identify synchronous events taking place at nearby positions, which is difficult with classical methods.