Hybrid dynamic classifier for drift-like fault diagnosis in a class of hybrid dynamic systems: Application to wind turbine converters

• Published in: Neurocomputing (Amsterdam) 2016-01, Vol.171, p.1496-1516
• Main Authors: Toubakh, Houari, Sayed-Mouchaweh, Moamar
• Format: Article
• Language: English
• Subjects: Data mining; Drift monitoring; Drift-like fault detection; Machine learning; Multicellular converters; Wind turbine
• ISSN: 0925-2312; 1872-8286
• DOI: 10.1016/j.neucom.2015.07.073

Hybrid dynamic systems (HDS) combine both discrete and continuous dynamics. Discretely controlled continuous systems (DCCS) is an important class of HDS in which the system switches between several discrete modes in response to discrete control events issued by a discrete controller. Their continuous dynamics depend on the discrete mode in which the system is. Wind turbine converters are an example of DCCS. Faults in converters may impact significantly the availability and the production performance of wind turbines. These faults can occur as a gradual abnormal change in the values of parameters describing the system continuous dynamics in a discrete mode. In this case, they entail a drift in the system operating conditions until the failure takes over completely. Detecting this drift in early stage allows reducing the power production losses as well as the wind turbine unavailability and maintenance costs. However, this drift can be observed only when the system is in the discrete modes where the continuous dynamics described by the affected parameters are active. Consequently, this paper proposes an approach based on the use of hybrid dynamic classifier able to monitor a drift in normal operating conditions of the converter in discrete modes where the continuous dynamics are impacted by a parametric fault. This allows keeping the useful patterns representative of the drift and therefore to detect it in its early stage.