Application of Feature Based Auto-Encoder for Anomaly Detection of Impulse Turbine in Wave Energy Convertor

초록

This study proposed a method for detecting abnormal motions of thrust bearings installed in impulse turbines, which are reciprocating rotary devices for power generation. The Anomalies were classified by applying an autoencoder, a type of deep learning algorithm, to the measured vibration data. Vibration data are collected during operation in the thrust bearing section of an impulse turbine. In order to analyze the characteristics of the measured vibration signals, statistical features of the signals were extracted at regular time intervals, and PCA (Principal Component Analysis) and machine learning-based classification algorithms were applied. The validity of the extraction section and the features for anomaly detection were confirmed. An autoencoder model for anomaly detection was created with three encoder and decoder layers, respectively, and the hyperparameters of the model were optimized by applying the grid search method. Afterwards, the residual of the signal predicted by the autoencoder and the actual measured signal was compared to calculate the anomaly detection threshold, which is the upper limit of anomaly detection, and an anomaly detection model and procedure were constructed. Finally, the measured test data was applied to the autoencoder failure detection model to confirm that the failure detection accuracy was over 99%.