On the Capacity of Deep Autoencoder-Based Normal Behaviour Models in Wind Turbine Condition Monitoring