Highly imbalanced fault diagnosis of turbine blade cracks via deep focal dynamically weighted conditional variational auto-encoder network

Turbine blade cracks pose a critical threat to the safety of aircraft in flight. However, data collected from turbine blades often exhibit substantial imbalances in real industry settings, posing challenges for effective blade crack diagnosis using intelligent models. To solve this problem, our study presents an intelligent fault diagnosis framework utilizing three-dimensional blade tip clearance (3D-BTC) signal under highly imbalanced samples, the so-called deep focal dynamically weighted conditional variational auto-encoder (DWCVAE-DFL), which can leverage the benefits of cost-sensitive learning methods and deep generative models to effectively tackles the diagnosis complexity of imbalanced class samples, thereby improving the blade crack diagnosis performance. Specifically, the novel deep focal loss (DFL) function is first designed to consider the varying sensitivity of imbalanced class samples in feature extraction and fault classification. The blade-rotor simulation rig is used for validating the effectiveness of our proposed method. Our experimental results indicate that the performance of our proposed method outperforms the latest methods when dealing with highly imbalanced samples.