Dual-head temporal modeling with causal normalization for cross-turbine gearbox fault early warning

Wind turbine gearboxes operate under fluctuating loads and widely varying conditions, which make cross-turbine generalization of supervisory control and data acquisition (SCADA)-based fault early warning difficult. In this article, a novel framework that combines uncertainty-aware residual modeling, dual-stream temporal representation, and causal amplitude normalization is introduced. Specifically, the raw oil-sump temperature is first mapped to a standardized residual through probabilistic regression, which mitigates load-dependent threshold drift. The residual is then routed into a dual-head architecture that models shift-type (mean-elevation) and compress-type (variance-contraction) degradation separately, which suppresses the false-alarm surge caused by conflating the two mechanisms. Finally, a four-layer progressive alarm engine with causal rolling p99 normalization and a training-free gating controller generates episode-level warnings without any target-domain retraining. The key innovation lies in combining mechanism-decoupled risk heads with self-referential normalization, which together enable direct deployment on unseen turbines of the same wind farm and turbine platform without fault-type prior knowledge. Evaluation on 2 years of SCADA data from five 2 MW onshore turbines, including a zero-shot blind-test turbine, shows that the deployable configuration achieves an early warning time of 39.2 h, a false alarm rate of 0.18 episodes/day, and an event recall of 1, demonstrating more stable cross-turbine generalization than the evaluated fixed-threshold and single-head baselines.