Early detection of internal erosion in earth dams: combining seismic monitoring and convolutional AutoEncoders

ABSTRACTLevees/earth dams are critical infrastructures for supplementing clean water, flood management, and energy production, prone to progressive failures due to internal erosion. Current inspection methods are unable to detect internal erosion until its exterior manifestation when it is too late to prevent the often-catastrophic failures. Therefore, finding innovative methods for the early detection of internal erosion is crucial. Despite the knowledge about the general mechanism of internal erosion, its early detection (and prevention) has remained a gap. This study introduces a novel artificial intelligence (AI) method to identify the temporal patterns within the passive seismic monitoring data, which can be associated with internal erosion initiation in earth dams. The proposed approach implements Convolutional AutoEncoders, an emerging deep-learning algorithm for anomaly detection in time-series data. Through an unsupervised learning framework, the AutoEncoders are trained using passive seismic monitoring data collected from a full-scale test embankment. In addition to the approximate initiation time, this algorithm can evaluate the initiation location by identifying the first sensors demonstrating internal erosion signs. The proposed deep learning framework combined with continuous seismic monitoring can serve as a basis for developing advanced early warning systems for internal erosion in earth dams.KEYWORDS: Internal erosionPassive seismic dataAnomaly detectionConvolutional autoencoderArtificial intelligenceEarth dams AcknowledgmentsThe authors would like to thank Dr Justin B. Rittgers (USBR) for providing the data and related support in this research and Dr Parisa Rahimzadeh Oskooei for her assistance in the data collection efforts. Grant funding for this research was provided by the University of Melbourne's Faculty of Engineering and IT (Early Career Research Grant held by Dr Negin Yousefpour).Disclosure statementNo potential conflict of interest was reported by the author(s).Additional informationFundingThis work was supported by University of Melbourne's Faculty of Engineering and IT (Early Career Research Grant): [Grant Number 1111].