Robust virtual sensing of turbine exit temperature for aero-engines using a truncated generalized correntropy-based extreme learning machine

Turbine exit total temperature ( T t6 ) is a critical yet challenging parameter to measure directly in aero-engines due to extreme thermal conditions that often lead to sensor faults like drift, threatening engine safety and efficiency. Virtual sensing provides a viable alternative, but prevailing data-driven methods, typically reliant on the mean squared error (MSE) loss, suffer from performance degradation under non-Gaussian noise and sensor anomalies. To overcome this limitation, this paper introduces a robust virtual sensing framework based on a novel Truncated Generalized Correntropy loss. By integrating the generalized maximum correntropy criterion with an adaptive truncation mechanism, the proposed loss function effectively suppresses the influence of outliers and faulty measurements. Embedded into an Extreme Learning Machine (ELM), this yields the Robust Truncated Generalized Correntropy ELM (RTGC-ELM) algorithm. The framework was rigorously validated using high-fidelity component-level model data under realistic flight profiles and further tested on the public NASA C-MAPSS dataset. Evaluations covered both normal operations and severe sensor fault scenarios (step-type and ramp-type drift). The results demonstrate that RTGC-ELM maintains high accuracy under normal conditions (MAE ∼1.21%) while exhibiting exceptional robustness under faults. For instance, under a step-type fault, RTGC-ELM limited performance degradation to only 0.01% (MAE increase from 1.21% to 1.22%), significantly outperforming conventional ELM (MAE increase from 1.28% to 1.75%) and other algorithms. This superior robustness was consistent across fault types and severity levels, confirmed through statistical significance tests and cross-dataset validation. The proposed RTGC-ELM provides a robust, efficient, and practical solution for analytical redundancy in aero-engine health management systems.