Lean burn combustion monitoring strategy based on data modelling

New designs of gas turbine lean burn combustors are under development to deliver lower emissions. To identify deterioration of combustion performance and engine health due to the increased complexity in these lean burn fuel system, one solution is through monitoring variation in Turbine Gas Temperature (TGT) profile. In this work, a data-driven monitoring strategy is designed and a prediction model for TGT associated with other crucial parameters is constructed. Due to limitations on sensing techniques and constraints on weight, only a limited number of TGT measurements downstream of combustion system are feasible in production engine, this along with gas swirling effects through the turbine, reduces the magnitude of temperature anomaly caused by an incipient fault. The model must meet EHM requirements on accuracy and sensitivity of the TGT monitoring model, be robust to influence of environmental changes. To accommodate these requirements, an adaptive model structure is proposed. A data-driven modelling framework with complexity control strategies for both a linear and a non-linear model are developed. The risk of overfitting is controlled by hyper-parameter optimization and cross-validation. The models are trained using data collected from combustor rig tests and test bed experiments. The fault mode behaviour is validated by augmenting the rig data with computational models of fault behaviour. Results show that with suitably selected range of data, and the application of the presented modelling framework, that a linear in parameter model provides an effective monitoring solution for lean burn systems. The adaptive modelling framework presented is also applicable to general data modelling tasks.