Annonce

Predictive maintenance, essential for the industry 4.0, allows for the prevention of costly breakdowns by anticipating
malfunctions. One way of planning maintenance operations is by detecting anomalies based on historical and realtime
data. However, the scarcity of data related to anomalies limits the effectiveness of supervised approaches, highlighting
the relevance of unsupervised methods. Time-frequency analysis [2, 5], which decomposes non-stationary
signals into temporal and spectral components, provides the opportunity to detect subtle variations in industrial systems.
The objective of this internship is to leverage time-frequency characteristics to formulate a regularized optimal
transport problem, leading to the development of an unsupervised anomaly detection algorithm. Indeed, optimal
transport [1, 6] is known for its sensitivity to anomalies, and the idea is to exploit this sensitivity in the time-frequency
domain to identify the types of considered anomalies [4]. The data for this study will be generated from a real industrial
environment using the IT’mFactory platform[3].
To implement our unsupervised algorithm for anomaly detection, we will focus on the following tasks:
• Modeling the behavior of coefficients in the time-frequency plane.
• Mathematical formulation of the anomaly detection as an optimization problem.
• Definition of a decision threshold based on optimal transport distance.