A steam turbine anomaly detection method based on O-DAE and SVDD

Abstract

Anomaly detection in unlabeled and highly imbalanced monitoring data is one of the most urgent to be solved and challenging industry problems. The use of autoencoders for anomaly detection is becoming more and more popular due to the powerful high-dimensional data analysis capabilities of autoencoders. A new anomaly detection method is developed base on O-DAE（optimized deep autoencoder） and SVDD（support vector data description）. Firstly， to make the training model hardly learns the features of abnormal samples， a sample screening mechanism is established to remove abnormal samples in the unlabeled training set. Secondly， the hidden features and reconstruction errors of the autoencoder are used as the final feature data for anomaly detection. Finally， the deep learning methods with different architectures are studied and compared， the experiments on the actual operation data of a steam turbine are conducted， and the detection abnormity is described combining with the support vector data. Compared with the traditional anomaly detection method， the anomaly detection accuracy of this method is improved by 50%， which can realize more sensitive and robust unsupervised anomaly detection of equipment performance for steam turbines.

Keywords

• deep autoencoder
• support vector data description
• steam turbine
• anomaly detection