Predictive maintenance model-based on multi-stage neural network systems for wind turbines

To optimise their maintenance schedules, reduce maintenance costs and extend the lifespan of their systems, wind farm operators are adopting new methods for predicting failures in advance. In this context, this paper proposes a multi-stage neural network model based on an Autoencoder and regression models to early detect the occurrence of an anomaly and predict the remaining useful life of a wind turbine in operation. Thus, we use regression and recursive techniques to adapt iteratively the remaining useful life model parameters. In this way, the model can gain knowledge from new weight inputs without losing the knowledge gained from previous inputs by maximising the accuracy. Meanwhile, we enrich the remaining lifetime model with the knowledge gained from the degradation dynamics. In other terms, we pay particular attention to the new insights without losing the knowledge gained from the previous one, thus maximising the global model accuracy. The methodology is applied on a wind farm dataset, provided by Energias De Portugal. Quantitatively, the process can improve the remaining useful life prediction accuracy by up to 16% compared to long-short memory (LSTM) and multi-layer perceptron (MLP) models.