Smart anomaly detection for Slocum underwater gliders with a variational autoencoder with long short-term memory networks

Abstract

Autonomous underwater vehicles (AUVs) are used extensively for monitoring the world’s oceans, taking measurements of oceanographic characteristics along the water column. Presently, there is no holistic anomaly detection system in operation and AUVs require experienced pilots to monitor the progress of missions. This results in a large operational overhead and reduces the number of AUVs that can be deployed simultaneously. This article proposes an online anomaly detection system for underwater gliders based on a data-driven approach. A novel Long Short-Term Memory (LSTM) Variational Autoencoder (VAE) has been developed and trained using field data from four deployments with healthy glider behaviour and then tested against four deployments where faults are present. The system is able to detect wing loss with a high degree of accuracy on gliders unseen during the models training, highlighting the generality of the model to different platforms. Additionally, the VAE method outperforms model-based solution for the detection of biofouling, proving its generality to different types of anomalies. The proposed smart anomaly detection will contribute to increasing the capacity of AUVs and reducing the dependence on support vessels and experienced pilots.