Fault Detection Relevant Modeling of an Industrial Gas Turbine based on Neuro-Fuzzy Approach

Analytical redundancy based fault detection and diagnosis system for an industrial gas turbine requires accurate model of the unit. Developing such a model from basic laws is a complicated and time consuming task. Besides, design parameters are difficult to get if the unit is already in operation. Parametric identification techniques are believed to provide substitute models. A reference model is compared with the real-time data to detect bias or incipient faults. A neuro-fuzzy (NF) reference model has been developed for the compressor discharge pressure and output electricity of an industrial gas turbine. Structure and network weights for the NF model are determined by a synergetic approach – Data clustering and Gradient Descent algorithm. Operation data collected in 10 seconds interval and for one day is used for model training and validation. Uncertainties are estimated assuming Gaussian error distribution and statistical independence. But, instead of considering all model parameters to define the lower and upper bounds, a test have been conducted if considering solely the linear part - consequent part of NF – can be a good substitute. The NF model is compared with the conventional radial basis function neural network model. It is shown that by NF model one can provide an accurate nonlinear description. And, with the developed confidence intervals included it is possible to detect abnormal states of the plant.