Highly imbalanced fault diagnosis of gas turbines via clustering-based downsampling and deep siamese self-attention network

For highly reliable gas turbines that rarely suffer faults, the overwhelming majority of historical data are collected under healthy state, while only a very small number of them are fault samples. However, traditional deep neural networks pay most attention to normal samples, resulting in a high missing diagnostic rate for fault samples. To address this problem, this paper develops a new fault diagnosis framework that integrates clustering-based downsampling with deep Siamese self-attention network (CBU-DSSAN), to reduce the number of normal training samples via clustering and strengthen the ability of fault feature extraction via multi-head self-attention mechanism. First, clustering-based downsampling is only conducted on the normal samples, and the cluster centers are put together with the fault samples to serve as the training data set, in order to balance the normal and fault classes. Second, the Siamese network maps the original data set into an embedded feature space, in which fault samples and normal samples belonging to different classes are far away from each other. Finally, the performance of the developed CBU-DSSAN has been evaluated using actual monitoring data of gas turbines.