Merging residuals from a set of simple normal behavior models to detect faults in wind farms

A fault detection method for wind turbines using Supervisory Control and Data Acquisition (SCADA) data is presented. The proposed approach generates seven residuals from simple normal behavior models predicting the temperature of the same turbine component and merges them to build faults detectors. The novelty of the proposed approach lies in the automatic creation of seven linear models with three input variables, which are identical for all wind turbines in a wind farm. The models are generated automatically from 1 year of fault-free SCADA data. The method is fully data-driven, has no parameters to tune, requires no expert knowledge, and can be applied to any wind farm and to any component whose failure causes a temperature increase. The use of simple models provides interpretable fault indicators for maintenance operators, while combining several models improves robustness to missing SCADA data. Both of these points are important industrial requirements. The method is validated on three wind farms comprising 22 wind turbines of three different technologies monitored over 1 year each, including three real fault cases. The results obtained confirm the feasibility of the proposed approach to assist operators responsible for the monitoring and maintenance of wind farms.