Condition monitoring is essential for ensuring the reliable operation and effective maintenance of modern complex systems, where nonlinear behaviors, environmental variability, and sensor noise can complicate fault diagnosis. Wind turbines, as representative examples of such systems, have widely adopted power curve-based methods for performance assessment and anomaly detection. However, most existing approaches overlook the heteroscedastic nature of noise, which adds uncertainty to model training and condition monitoring. In this paper, we propose a Bi-level Piecewise Linear Model (Bi-PLM) with physics-informed constraints to improve resilience against potential data contamination. A data-driven procedure, combining a binning method with locally estimated scatterplot smoothing (LOESS), is developed to characterize the noise heteroscedastic structure, which is then incorporated into change point detection (CPD)-based condition monitoring. Experiments on two real-world datasets show that explicitly accounting for heteroscedasticity reduces variance-induced uncertainty in residuals and substantially lowers false positives in fault detection, yielding average increases in the area under the curve (AUC) of approximately 6.0% and 8.4% on the two datasets. Comparative results against benchmark models confirm the robustness and reliability of the proposed method.
Gao et al. (2026) studied this question.