Anomalies In The Depths: Evaluating Unsupervised Anomaly Detection on Real-World Unlabelled Data. A Case Study on Autonomous Ocean Gliders

Unsupervised anomaly detection (AD) on scientific sensor data is challenging due to the absence of labels, heterogeneous sequence lengths, and mixed anomaly types. We present a case study and an evaluation process for AD on multivariate data-sequences recorded by an autonomous ocean glider. The dataset consists of 341 variable-depth flights (3-920 m, 1 m vertical resolution) collected by glider IFM03 during R/V Meteor's cruises M105 and M106 in the Central Tropical Atlantic. We train a compact variational recurrent autoencoder with a CNN preprocessing layer and bidirectional LSTM encoder (16-dimensional latent) and a lightweight decoder. Reconstruction error (MSE) yield flight-level anomaly scores. In lieu of labels, we evaluate the AD system by means of oceanographic knowledge via expert inspection, separation of reconstruction error between normal flights and anomalies, and clustering of the learned embedding. Moreover, we present the result of the model in its geospatial context. Our contribution is a practical case-study of deploying unsupervised AD on real-world, unlabelled oceanographic data.