INTELLIGENT FAULT DETECTION AND DIAGNOSIS SYSTEM FOR A COGENERATION AND COOLING PLANT

Recent designs of Cogeneration and Cooling Plants (CCPs) are getting more complex, with the addition of a variety of mechanisms to increase overall system efficiency
while keeping COz and NO, emissions as minimum as possible. Meeting the two needs over the plant life time is difficult unless suitable provisions are included to
deal with abnormal conditions that may be caused by aging, malfunction, wrong calibration of sensors, etc. Maintenance cost of power producing plants may reach up
to about 30% of the total power generating cost. Recent studies also indicate that significant reduction of the cost may be achieved by shifting from preventive
maintenance to Condition Based Maintenance (CBM). The main part of CBM is a suitable Fault Detection and Diagnosis (FDD) system. Because CCPs often have high
energy throughputs, any attempt to improve the existing FDD approaches is linked to high economic incentives. CCPs by nature exhibit considerable nonlinearity in their
dynamics. Due to this, condition monitoring based on linear models may lead to poor results. CCPs are also featured by multiple operating regions, making fault detection
and diagnosis a difficult task. During operation, a fault may occur in one of the regions. The volume of data is also another problem. Therefore, the FDD system
design must take into account the stated factors. The objective of this thesis is to develop a fault detection and diagnosis system based on nonlinear models that use
computational intelligence techniques.
The unavailability of many design point information for modeling and the need to design a fast FDD system necessitated the reliance on data based methods. To capture
trends for normal operating conditions, models are developed by applying neurofuzzy
method. For the dynamic case, the same approach is used in the framework of generalized orthonormal basis functions. Since using fixed model threshold in the
fault detector could cause false alarms or missed fault detection, the proposed method
uses adaptive techniques for the estimation of model uncertainty.