Adaptive training of a vibration-based anomaly detector forwind turbine condition monitoring system (CMS) is carriedout to achieve high-performance detection from the earlystages of monitoring. Machine learning-based wind turbineCMSs are required to collect large-scale data to yield reliablepredictions. Existing studies in this area have postulatedthat both data for training a monitoring system and those duringthe operation of the system are obtained from identicaldevices. In addition, constant monitoring of data is desirable,but in practice, the data can be observed periodically(e.g., several tens of seconds of data are observed every twohours). In this case, collecting sufficient data is time consuming,making it difficult to conduct accurate predictions atthe early stage of the CMS operation. To address this problem,a small amount of vibration data observed at a targetwind turbine is utilized to adapt the anomaly detector thatis trained on relatively large-scale vibration signals obtainedfrom other wind turbines. In the present study, maximum aposteriori (MAP) adaptation is applied to a Gaussian mixturemodel (GMM)-based anomaly detector. Experimentalcomparisons using vibration data from the gearbox in the experimentalenvironment and those used in the wind turbinedemonstrated that MAP-based GMM adaptation yielded animprovement in anomaly detection accuracy even when onlya small amount of data is observed at the target gearbox.