Gas turbine thrust estimation in sensor drift scenarios using a three-stage multi-target domain adaptation method

To mitigate the impact of single sensor drift on gas turbine thrust estimation and enhance fault tolerance, this study introduces a three-stage multi-target domain adaptation (TSMTDA) approach. Initially, the feature selection stage employs a heterogeneous ensemble feature selection strategy combining wrapper and filter techniques to select an input feature set with low sensitivity to drift, reducing distribution disparities between source and target domains from the feature level. Subsequently, in the dataset construction stage, the salp swarm algorithm is utilized to optimize the weight coefficients across different domain datasets, promoting accurate multi-domain data learning while reducing training time. Finally, in the algorithm design stage, a knowledge distillation-based multi-target domain adaptation extreme learning machine is developed for method formulation, ensuring the adaptability to arbitrary sensor drift from the algorithmic level. Comparative experiments validate the effectiveness of the TSMTDA approach in mitigating the impact of sensor drift, with the TSMTDA-based thrust estimation method demonstrating superior performance. Even with sensor drifts up to ±10 %, the maximum relative error remains as low as 3.411 %. Furthermore, ground tests conducted on a micro turbojet engine further demonstrate the method's robust engineering applicability.