Acoustic Events Detection and Classification for Urban DAS Data with Supervised Learning

Abstract A key aspect of improving disaster prevention and mitigation in sustainable smart cities is to increase the sensory capabilities of existing communication infrastructure, providing reliable information for urban management in emergency situations. Distributed acoustic sensing (DAS) is an advanced technology suitable for this application because it has a wide range of applications, including urban environmental awareness, structural health monitoring, and disaster warning. In this study, the field test data are measured by a DAS array deployed along the edge of the Guye area in the city of Tangshan in China, where the 1976 Tangshan earthquake occurred. We analyzed the vibrations from natural and artificial acoustic sources across both the space and frequency domains and revealed various characteristics of the sources. Subsequently, a deep learning-based method was developed for multiple acoustic source detection and classification, including earthquake, vibrator vehicle, traffic flow, and industrial production. The training dataset was created using this acquisition of DAS field data, which was annotated using the label transfer method proposed in this article. Then, typical acoustic events are classified and extracted from DAS data in the space–frequency domain. The proposed source identification scheme enables real-time monitoring of routine urban activities with long-distance coverage and high accuracy, as well as detection of abnormal events. In addition, we can use this method to expand the range of recognized classes or apply it to special datasets. Our study shows a great value in improving the ability of urban environmental perception and hazard information analysis. It also holds potential for earthquake detection, site-effects studies, and anomaly detection in urban environments.