A Cost-Aware Deep Learning Framework for Gearbox Fault Detection in Wind Turbines

The maintenance operations in the wind sector (particularly offshore) are extremely costly, and thus technologies for optimization of maintenance need to take the high risks related to these costs into account. On the topic of predictive maintenance, any false prediction can lead to high cost incursions, and even a few false predictions can invalidate the predictive maintenance strategy. The costs and benefits of correct and false predictions are not uniform or symmetric, leading the conventional performance metrics to misalign the perceived model performance with the actual business impact. This problem leads us to introduce a cost-aware deep learning framework that accounts for the costs and benefits of predictions. This framework is achieved by loss functions designed to adjust the parameters according to the cost–benefit and not to optimize model accuracy. For the case study of wind turbine generator gearbox failure prediction, several cost-aware loss functions are tested, along with a novel fine-tuning approach that combines them. The case study shows that using cost-aware expected-value loss functions improves cost–benefit performance by an average of 30–57% compared with models trained with conventional loss functions.