AN INTEGRATED SCHEME FOR ADAPTATION AND UPDATING OF ANOMALY DETECTION MODEL
Presently, most anomaly detection learning algorithms assume that the training data is purely normal and could be directly compared with the testing data. However, two problems are usually encountered in the processes of application: firstly, training data might contain a number of anomalies due to the sensor malfunction or environmental effect, so that the model could not be expected to respond to the similar anomalies appeared in the testing data; Secondly, the trained model with the historical training data may have nothing to do with the new coming events, because of the changes of speed range or load regime. To address these two issues, in this paper, the novel approaches have been developed based on Gaussian mixture models (GMMs). At the beginning, the entropy score and distance based methods are developed for the trained Gaussian components to ensure the trained model represents the real normal condition; next, a new procedure is introduced to identify appropriate background level for the testing data before the data are utilized to update anomaly detection model through transferring and merging processes. Finally, synthetic and experimental bearing data are used to validate these approaches.
Anomaly detection Gaussian mixture model Training data.
S.L. Chen R.J.K. Wood L. Wang R. Callan H.E.G. Powrie
The national Centre for Advanced Tribology at Southampton (nCATS), School of Engineering Sciences, U GE Aviation, Digital Systems, Chandlers Ford, SO 53 4YG, UK
国际会议
北京
英文
284-296
2008-10-27（万方平台首次上网日期，不代表论文的发表时间）