An adaptive on-board real-time model with residual online learning for gas turbine engines using adaptive memory online sequential extreme learning machine

Abstract

The on-board real-time model (ORM) of gas turbine engines (GTEs) is widely used in various applications of control systems, such as sensor fault-tolerant control and model-based control. However, the accuracy of the ORM in the primary stage is affected by changes in engine performance, including degradation and sudden changes, which inevitably occur during engine service. To address this issue, we propose an adaptive on-board real-time model (AORM) modelling framework for GTEs based on a component-level model (CLM) and a residual online learning model (ROLM). The CLM is constructed using classical component-level modelling theory, while the ROLM is built using the proposed adaptive memory online sequence extreme learning machine (AMOSELM) algorithm. The AMOSELM algorithm can effectively learn the residuals between the engine and the CLM during engine service, improving the accuracy of the model. Specifically, we compensate for the CLM by using the predicted residuals based on the AMOSELM to achieve enhanced accuracy in the AORM. The memory coefficient of the AMOSELM is adjusted using a bell-type membership function in the online learning stage. The shape parameters of the membership function are updated using an off-line optimization strategy based on particle swarm optimization when the online learning is finished. To validate the effectiveness of the AORM framework and the AMOSELM algorithm, we conducted both virtual engine flight data simulation and actual ground test verification. The verification results show that the AMOSELM algorithm can effectively realize residual online learning during engine service, and performs better than other traditional online learning algorithms. Moreover, using the predicted residuals to compensate for the CLM can significantly reduce the modelling error of the AORM, achieving better modelling accuracy.