A novel approach to gas turbine fault diagnosis based on learning of fault characteristic maps using hybrid residual compensation extreme learning machine-growing neural gas model

Gas path analysis is a well-known approach to gas turbine condition monitoring that consists of three steps, including detection, isolation and identification of performance deteriorations. Commonly, these steps are performed consecutively, and individual algorithms are exploited for each step, which in turn adds to the complexity of the diagnostic system. To tackle this problem, this paper proposes a novel gas turbine fault diagnosis approach to simultaneously detect, isolate and identify the gas path faults. This approach is based on learning the fault characteristic maps (FCMs) of gas turbine components using the growing neural gas (GNG) network and residual compensation extreme learning machine (RCELM). First, a bank of RCELMs is trained to estimate the health parameter vector. The GNG network is then used as a tool to learn the topology of the maps (FCMs) in a manner that each neuron of the network represents a potential health condition of the gas turbine along with a certain deterioration severity. Since the position of each neuron on the map indicates a specific health state, the GNG network is able to map the output of RCELMs (the health parameter vector) to a certain health condition, resulting in fault detection, isolation and identification of the monitored components of the engine all together. The performance of the proposed approach is assessed against data collected from a two-shaft industrial 25 MW gas turbine model, and it is demonstrated that the proposed diagnostic tool is capable of fast and accurate detection, isolation and identification of anomalies in the main components of the engine.