We propose a SCADA data-based method for fault detection on wind turbines, based on an ensemble of linear indicators. In a previous work ([1]), a model generation process using multi-turbine residual indicator has been developed. This process uses three-variable linear normal behavior models. While results have shown good performance detection on real cases, some limitations were observed: the false alarm rate remains high on fault-free periods, and some faults remains undetected.

The significative false positive rate can be explained by the linear modeling approach, and the low frequency of data acquisition used. Concerning the non-detection, it can be explained by the fact that the model generation process is based performances on healthy period, thus it may selects regressors that are correlated during a faulty period.

In both cases, using a larger number of variables for the monitoring of a component may be a way to solve the problem. In this study, we propose a method to automatically generate a set of linear models, predicting the evolution of a same variable. The process is based on a greedy forward selection algorithm, generating five unique linear models. Each model is made different by forcing the variable selection algorithm to use different input variables. An additional physical model is also considered. Multi-turbine health indicators are then built using these models, together with a signal comparing the variation of the variable to be estimated at the wind farm level.

This study present a comparative analysis of the detection performance of the initial single-indicator monitoring approach and the one using the average residual indicator. The method is evaluated on two real fault cases. The results show that the multi-indicator monitoring approach overcome the limitations of the single-indicator one on healthy and faulty period.