A State-Supervised Model and Novel Anomaly Index for Gas Turbines Blade Fault Detection Under Multi-Operating Conditions

To meet industrial demands, gas turbines typically operate under multiple conditions, presenting unique challenges for fault diagnosis. This paper proposes a novel blade fault detection framework designed for such environments. First, a State-Supervised Variational Autoencoder (SS-VAE) model is introduced, which integrates the learning process of turbine operational states into the VAE bypass, enabling it to capture variations in vibration signal data across different operating conditions. Through the fusion decoding of operating states and vibration signals, the foundation of the SS-VAE model and its loss function is established. Secondly, a new State Mapping Error (SME) index is introduced to further address the issue of reconstruction error variability across multiple conditions. The proposed method is validated through a blade fault test rig and applied in an industrial gas turbine blade fault case. Results demonstrate that the SS-VAE model and SME index effectively capture vibration signal changes due to blade faults across multi-operating conditions, achieving accurate monitoring and diagnosis. Compared to traditional methods, the proposed approach shows higher effectiveness and robustness.