Cost-optimized probabilistic maintenance for condition monitoring of wind turbines with rare failures

We propose a method, a model, and a form of presenting model results for condition monitoring of a small set of wind turbines with rare failures. The main new ingredient of the method is to sample failure thresholds according to the profit they give to an operating company. The model is a multiple linear regression with seasonal components and external regressors, representing all sensor components except for the considered one. To overcome the scarcity of the training data, we use the median sensor values from all available turbines in their healthy state. The cumulated deviation from the normal behavior model obtained for this median turbine is calibrated for each turbine at the beginning of the test period and after known failures. The proposed form of presenting results is to set a scale for possible costs, control for random maintenance, and show a whole distribution of costs depending on the free model parameters. We make a case study on an open dataset with SCADA data from multiple sensors and show that considering the influence of turbine components is more critical than seasonality. The distribution, the average, and the standard deviation of maintenance costs can be very different for similar minimal costs. Random maintenance can be more profitable than reactive maintenance and other approaches. Our predictive maintenance model outperforms random maintenance and competitors for the whole set of considered turbines, giving substantial savings.