Proceedings of the

The 33rd European Safety and Reliability Conference (ESREL 2023)
3 – 8 September 2023, Southampton, UK

On the Joint Use of An Ensemble of Linear Residuals to Improve Fault Detection in Wind Turbines

Théodore Raymond1,a, Sylvie Charbonnier2,c, Alexis Lebranchu1,b and Christophe Bérenguer2,d

1Data Engineering, Valemo, France.

2Université Grenoble Alpes, CNRS, Grenoble-INP, GIPSA-Lab, France.


One of the biggest levers for reducing the cost of wind power generation is to minimize the replacement frequency of large components. To address this need, researchers have focused on the development of real-time health monitoring of component to perform condition-based maintenance. In a previous work, a fault detection solution based on multiturbine indicators built from automatically generated linear models has been presented and validated on a converter fault case. However, the application of this method on other faults revealed weaknesses in the detection performance, making the solution unreliable. To address these issues, the solution proposed in this study is to consider an ensemble method to automatically generate a set of tri-variable linear models predicting the evolution of a common variable. The linear models are constructed using a constrained greedy selection algorithm, providing unique sets of model variables. From these models, residual-based multi-turbine health indicators are constructed, and a mean linear residual is considered, computed as the mean of seven different indicators. The comparative analysis of these indicators, carried out based on the area under two receiver operating characteristic curves on two fault cases, shows that the use of a mean linear residual computed from a set of linear residuals significantly improves the global detection performance, and thus the reliability of the condition-based maintenance process under development.

Keywords: Wind energy, Linear multi-variable models, Ensemble methods, Residual-based fault detection, Intelligent variable selection, Automatic model generation.

Download PDF