Deep Learning for fault detection in wind turbines. (December 2018)

• Record Type: Journal Article
• Authors: Helbing, Georg; Ritter, Matthias
• Is Part Of: Renewable & sustainable energy reviews. Volume 98(2018)
• Page Start: 189
• Page End: 198
• Publication Date: 2018-12
• Subjects: ANN Artificial Neural Network -- API Application Programming Interface -- AUC Area Under the Curve -- CNN Convolutional Neural Network -- EWMA Exponentially-Weighted Moving Average -- GPU Graphical Processing Unit -- GRU Gated Recurrent Unit -- LSTM Long-Short Term Memory -- MD Mahalanobis Distance -- MLP Multi-Layer Perceptron -- NBM Normal Behaviour Model -- PCA Principal Component Analysis -- RBM Restricted Boltzmann Machine -- RNN Recurrent Neural Network -- ROC Receiver Operating Characteristics -- SAE Stacked Autoencoder -- SCADA Supervisory Control and Data Acquisition -- SDAE Stacked Denoising Autoencoder -- SMLDAE Stacked Multilevel-Denoising Autoencoder -- SPC Statistical Process Control; Deep Learning -- Artificial neural network -- Fault detection -- Condition monitoring -- Wind turbine; Renewable energy sources -- Periodicals; Power resources -- Periodicals; Énergies renouvelables -- Périodiques; Ressources énergétiques -- Périodiques; 333.794
• DOI: 10.1016/j.rser.2018.09.012
• Languages: English
• ISSNs: 1364-0321

Abstract

Abstract: Condition monitoring in wind turbines aims at detecting incipient faults at an early stage to improve maintenance. Artificial neural networks are a tool from machine learning that is frequently used for this purpose. Deep Learning is a machine learning paradigm based on deep neural networks that has shown great success at various applications over recent years. In this paper, we review unsupervised and supervised applications of artificial neural networks and in particular of Deep Learning to condition monitoring in wind turbines. We find that – despite a promising performance of supervised methods – unsupervised approaches are prevalent in the literature. To explain this phenomenon, we discuss a range of issues related to obtaining labelled data sets for supervised training, namely quality and access as well as labelling and class imbalance of operational data. Furthermore, we find that the application of Deep Learning to SCADA data is impeded by their relatively low dimensionality, and we suggest ways of working with higher-dimensional SCADA data. Highlights: An overview of recent applications of artificial neural networks and Deep Leaning. Most approaches in the literature are unsupervised. Supervised methods that use high-dimensional input data are quite successful. Supervised approaches are impeded by issues with operational data. These regard quality, availability, dimensionality, labels, and class imbalance.