State of the Art : Classification algorithms for sensor fusion in the approach phase Cocotier Project
Résumé
This report first formulates the problem that must be approached in the work package LOT 5.9 as well as the identified challenges to solve this problem with a data driven approach (model based approaches are the focus of another work package). It then presents a state of the art of the methods that are commonly used to solve a fault detection problem and it provides an overview of the different data driven approaches applicable to the problem.
Standard unsupervised learning methods (Section 3.1) as well as supervised learning methods (Section 3.2) are presented. A focus is then put on data driven dynamic methods (Section 3.3) that allow one to account for dynamics and evolving environments, in par- ticular dynamic clustering methods are presented in Section 3.3.1, including the in-house dynamic clustering tool DyClee, and correlation-based methods are presented in 3.3.2.
Specific methods that can be used to face identified challenges (fusion, imbalanced data) are introduced in Section 4 and some metrics that can be used to evaluate a model obtained from a learning process are given.