Sparse spatiotemporal feature learning for pipeline anomaly detection

Abstract

Spatiotemporal systems are often difficult to represent in the presence of noise and exhibit higher modelling complexity. This scenario is found in infrastructure applications such as continuous pipeline monitoring, where it is of interest to pinpoint abnormal situations. Non-stationarity in the pipe response to changes in flow also complicates the detection. To determine anomalous events in the pipe, a framework is developed where pipe dynamics are modelled through fiber-optic acoustic measurements during pipe flow, and deviations in the predicted data are flagged for anomalies. An approach based on state space embedding of pipeline dynamics behaviour is developed. Sparse feature learning using autoencoders encodes the state space for detecting events and predicting pipe acoustic behaviour. Results based on experimental data show the effectiveness for pipeline monitoring in the presence of additive noise and spatial information.