Fault detection and isolation method for gas turbines using self-organizing type-3 fuzzy wavelet neural networks

Area of Science:

• Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Data-driven fault detection and isolation (FDI) in gas turbines is crucial for operational efficiency and safety.
• Existing FDI methods are often brittle, lacking adaptability and robustness to noise and uncertainty.

Purpose of the Study:

• To develop a self-organizing, adaptable, and robust FDI system for gas turbines.
• To enhance uncertainty handling and model compactness in FDI schemes.

Main Methods:

• Introduction of a Type-3 fuzzy membership function for improved uncertainty management.
• Hybrid optimization using Adam and Unscented Kalman Filter for neuro-fuzzy architecture training.
• Implementation of a self-organizing mechanism for adaptive rule generation and pruning.

Main Results:

• Achieved high average FDI rates: 99.302% for detection and 99.324% for isolation on a Siemens gas turbine simulator.
• Demonstrated superior performance over state-of-the-art fuzzy systems (FSRE-AdaTSK, TSK-SRB) with fewer rules.
• Showcased competitive performance against deep learning models (Transformers, LSTMs, CNNs) using significantly fewer parameters.

Conclusions:

• The proposed Self-organizing Type-3 Fuzzy Rough Wavelet Neural Network (ST3FRWNN) offers a practical and deployable solution for gas turbine health monitoring.
• ST3FRWNN exhibits excellent robustness to noise (20-dB SNR) and computational efficiency.
• The adaptive and self-organizing nature of ST3FRWNN ensures accurate and reliable FDI in dynamic operational environments.