The “Online Deep anomaly Detection” (ODD) Project

Summary

The ODD Project is a 4-year research project that addresss Online Deep anomaly Detection (ODD), funded by the French research agency (ANR) under the grant ANR-23-CE23-0004.

Anomaly detection is a challenge per se. It is unsupervised by nature, since abnormal events are rare, varied, and cumbersome to collect. By exploring deep neural networks for learning efficient representations, the major categories of deep anomaly detection methods are deep one-class classification, autoencoders, generative adversarial networks, as well as self-supervised learning.

The ODD Project aims to address several difficulties that arise in anomaly detection when dealing with time series. Time series have some specificities, such as concept drift, seasonality, nonstationarity, etc, in addition to several modalities; Anomaly highly depends on the temporal contextualization; and early detection should be made in real-time. To this end, three research directions are investigated in this project

  • We revision recent advances in anomaly detection in view of the online paradigm, with a focus on deriving meaningful deep representations of the temporal information in hierarchical and multiresolution models. 
  • We explore the optimal transport (OT) theory to provide online OT-based methods for detection and distribution shifts, with the false alarm minimization under an early detection strategy.
  • We revisit the statistical change-point detection theory in the light of deep neural networks, and more specifically generative models such as diffusion probabilistic models and normalizing flows.

While the major contributions will essentially be fundamental in Machine Learning, the project will address several applicative domains. Two application areas are of particular interest: complex multivariate time series for health condition monitoring of rotating machinery, and spatiotemporal data for pollution detection in air, soil, or water.

Principal investigators

  • Paul Honeine (40%)
  • Fannia Pacheco (30%)
  • Gilles Gasso (20%)
  • Laurent Heutte (20%)
  • Maxime Berar (20%)
  • Simon Bernard (20%)
  • Abdel Ennaji (20%)
  • Pierrick Tranouez (20%)
  • Benoît Gaüzère (10%)

Non-permanent members

  • Romain Mussard (PhD), working on label shift matching for anomaly detection and classification in time series
  • Marwan Taia (PhD)
  • Assmaa Al Samadi (PhD)
  • to be hired (PhD)
  • to be hired (Post-doc)