Using supervisory control and data acquisition (SCADA) data to detect faults in wind turbines (WTs) has gained interest over the last few years. The SCADA system is installed by default for modern WTs and a condition monitoring system can be employed without installing additional measurement devices, which ensures a cost-effective solution for operators. Most systems developed today monitor only one component at a time. To cover all aspects of aWT's operation one would therefore have to use one model for each component. Such a system would quickly become unwieldy and expensive to manage in practice. This paper proposes a model based on the autoencoder, a neural network that reconstructs all its input signals. The network is trained on healthy data and will therefore only give a good reconstruction on data which has the same characteristics. This model is capable of monitoring aWT holistically: a single model can detect failures in multiple components. A strategy for designing autoencoder models is described and various hyperparameters that affect the performance of the models are investigated. Finally, the best performing model is chosen and its performance as an anomaly detection tool is demonstrated with case studies from real turbine data. (C) 2020 Elsevier Ltd. All rights reserved.