Gas turbine failure classification using acoustic emissions with wavelet analysis and deep learning

Abstract

Compared to vibration monitoring, acoustic emission (AE) monitoring in gas turbines is highly sensitive to changes that do not involve whole-body motion, such as wear, rubbing, and fluid-induced faults. AE signals captured by suitably mounted sensors can potentially provide early indications of abnormal turbine operation before such abnormalities manifest in structural vibration or emitted airborne noise. However, developing an online fault detection system requires extensive real-time data treatment to extract appropriate features and indicators from raw AE records. To build such a system for industrial turbines, researchers need to understand the AE-generating mechanisms associated with turbine operation and the sources of background noise. In this study, we aim to develop such an understanding using a small-scale turbine whose operational conditions can be modified safely to reflect both normal and faulty conditions. Our signal processing approach involves first extracting a time-series envelope using an averaging time selected to enhance major features and eliminate irrelevant noise. We then generate time–frequency features using a continuous wavelet transform, which are used to train a deep convolutional neural network to classify gas turbine conditions. The resulting model demonstrates high accuracy in classifying two normal running conditions and two faulty conditions at various turbine speeds. Overall, the proposed methodology offers a powerful tool for gas turbine condition monitoring, and we make all associated data available in open-source format to facilitate further research in this field.4