Fleet‐Wide Data‐Driven Model Generation Process for Residual‐Based Fault Detection in Wind Energy
ABSTRACT In order to reduce curative interventions in wind energy, maintenance approaches based on real‐time monitoring of component health have been widely studied. These condition‐based maintenance processes generally rely on data‐driven residual indicators, measuring the difference between a meas...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Wiley
2025-09-01
|
| Series: | Wind Energy |
| Subjects: | |
| Online Access: | https://doi.org/10.1002/we.70050 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | ABSTRACT In order to reduce curative interventions in wind energy, maintenance approaches based on real‐time monitoring of component health have been widely studied. These condition‐based maintenance processes generally rely on data‐driven residual indicators, measuring the difference between a measured value and an estimation made by a normal behavior model. These solutions face limitations that prevent their applicability to an industrial context. First, models, which are mostly built using artificial neural networks, can be very accurate but remain difficult to interpret by a wind turbine operator and present implementation constraints due to the large number of inputs variables. Moreover, existing condition monitoring processes proposed for wind turbines are defined for a scope reduced to a single manufacturer and to a set of components that does not include all the critical ones, which does not guarantee their possible deployment across a fleet. In this paper, a data‐driven automatic model generation method for fault detection, applicable to an entire fleet of wind turbines, is presented. The models obtained are simple, physically coherent, and valid for all the wind turbines of a farm. Multiturbine residual indicators, robust to environmental variations, are then built and analyzed to perform fault detection. The fault detection ability of these indicators is evaluated on four faults impacting four components from different wind farms and manufacturers, using receiver operating characteristic curves. Finally, the fault detection performances of the proposed models are evaluated on wind farms with turbines of various technologies, whose data have not been used to learn the models. The good results obtained show that the implemented process is suitable for industrial deployment, and is of interest to companies wishing to optimize the profitability of their fleet. |
|---|---|
| ISSN: | 1095-4244 1099-1824 |