An innovative data-driven AI approach for detecting and isolating faults in gas turbines at power plants

This study investigated the detection and isolation of gas path faults in a power plant gas turbine using efficiency data and fundamental quantities. First, attention is given to balancing data and selecting instances. Two new neural-fuzzy networks were then designed and trained using the Hippopotamus optimization algorithm. Developing these two networks aims to create a network resilient to noise with high accuracy and a low parameter count. Third, a broad spectrum of Artificial Intelligence based methods, such as shallow neural networks, machine learning models, and deep learning models, were employed to compare the proposed networks for fault detection and isolation of one power plant 163 MW gas turbine from Siemens Company. The investigation results indicate that the proposed hierarchical structure achieved an average of 99.81 % for fault detection and 99.50 % for fault isolation, consisting of only 203 learning parameters for fault detection and 335 for fault isolation, and operates better than the methods mentioned above in terms of accuracy, precision, sensitivity, and F1-Score metrics criteria.