Wind turbines are susceptible to failure events which reduce operational availability, increase costs and introduce safety hazards. Anomaly detection can be used to identify failure events or provide early warnings of faults. Many industrial installations rely on fault detection methods based on static thresholds over parameters estimated from the condition monitoring data, following established standards. However, static thresholds may fail to capture the variability of operational contexts and the individual turbines’ behaviour. Adaptive machine learning methods offer instead data-driven adaptation flexibility. However, this flexibility is restricted by the narrow range of situations which the operational data are drawn from, mostly corresponding to normal operating states or a limited set of failure events. To address such challenges, this study combines a Variational Autoencoder with an Isolation Forest to effectively capture anomalies. The performance of this method is evaluated against static thresholds recommended in established standards over benchmarking data from the domain, as well as additional operational data from wind turbines. Results indicate that, while static thresholds may suffice in simpler scenarios, they often fail in more complex operational ones. In contrast, the proposed model demonstrates adaptation capacity and is more successful in detecting anomalies even with noisy and sparse data, indicating promising operational potential for industry.