Anomaly Detection in Gas Turbines Using DLSTM - Autoencoder with Data Augmentation

The present work aims to enhance fault detection in the Kirkuk power plant gas turbines based on a hybrid machine learning model. Autoencoder and a Deep Long Short-Term Memory (DLSTM) hybrid model was used for anomaly detection out of vibrational data. The dataset was collected through CA 202 piezoelectric accelerometers mounted on the turbines. Since there was little abnormal data, augmentation was employed to create synthetic anomalies for the dataset by flipping and applying jittering to the same dataset. The desired hybrid model achieved an anomaly detection accuracy of 96.10% for the Clark dataset and outperformed the Isolation Forest and K-means clustering. As a result, the present work offers a solid foundation to drive the predictive maintenance process for industrial gas turbines and thus reduce the likelihood of a failure occurrence.

Anomaly Detection in Gas Turbines Using DLSTM - Autoencoder with Data Augmentation Keywords:

Power plant , Gas turbine , Fault detection , Data augmentation , Predictive maintenance