Deep anomaly detection using self-supervised learning: application to time series of cellular data

Romain Bailly (1, 2) , Marielle Malfante (1) , Cédric Allier (3) , Lamya Ghenim (4) , Jérôme I. Mars (2)

1 LIIM - Laboratoire Intelligence Intégrée Multi-capteurs
2 GIPSA-SIGMAPHY - GIPSA - Signal Images Physique
3 CEA-LETI - Commissariat à l'énergie atomique et aux énergies alternatives - Laboratoire d'Electronique et de Technologie de l'Information
4 IRIG - Institut de Recherche Interdisciplinaire de Grenoble

We present a deep self-supervised method for anomaly detection on time series. We apply this methodology to detect anomalies from cellular time series. In particular, this study focuses on cell dry mass, obtained in the context of lensfree microscopy. The method we propose is an innovative two-step pipeline using self-supervised learning. As a first step, a representation of the time series is learned thanks to a 1D-convolutional neural network without any labels. Then, the learned representation is used to feed a threshold anomaly detector. This new self-supervised learning method is tested on an unlabelled dataset of 9100 time series of dry mass and succeeded in detecting abnormal time series with a precision of 96.6%.

Keywords: Self-supervised learning; 1D-CNN; Anomaly detection; Cellular anomaly; Time series; Lens-free microscopy

Domains: Engineering Sciences [physics]; Signal and Image processing; Computer Science [cs]; Artificial Intelligence [cs.AI]

Romain Bailly, Marielle Malfante, Cédric Allier, Lamya Ghenim, Jérôme I. Mars. Deep anomaly detection using self-supervised learning: application to time series of cellular data. ASPAI 2021 - 3rd International Conference on Advances in Signal Processing and Artificial Intelligence, Nov 2021, Porto, Portugal. ⟨cea-03605065v1⟩