Physics-Informed CNN–BiLSTM with Survival Analysis for Wind-Turbine SCADA Predictive Maintenance: A Controlled External Benchmark on Penmanshiel and Kelmarsh

Abstract

Wind-turbine Supervisory Control and Data Acquisition (SCADA) predictive maintenance is evaluated under operational class imbalance, distribution shift across turbines and sites, and right-censored failure horizons. A leakage- safe turbine-holdout split with split-wise preprocessing is applied with external validation on two public Senvion cohorts (Penmanshiel, primary; Kelmarsh, secondary). The learning core is a physics-informed Convolutional Neural Network– Bidirectional LSTM (CNN–BiLSTM). Temporal convolutions provide local inductive bias. A bidirectional long short- term memory (LSTM) captures longer-range context. A drivetrain torque-balance residual enters the training loss (not as engineered inputs) as a physics-regularised penalty term. Eleven Statistical Process Control (SPC) indicators are concatenated onto the learned embedding for operator auditability. A Cox proportional-hazards head provides censoring- aware Remaining Useful Life (RUL) estimates. Across ten matched seeds with bootstrap 95% confidence intervals, receiver operating characteristic area under the curve (ROC-AUC) on the headline RUL classification label reaches 0.932 ± 0.008 (Penmanshiel) and 0.993 ± 0.002 (Kelmarsh) under the full-channel evaluation protocol. On the matched 48 h RUL band, tree ensembles reach ≈ 0.60 ROC-AUC (Table 5); that comparison is directional (label, prevalence, and feature-width asymmetry noted in the table footnote). Paired Wilcoxon tests against tree baselines under each cohort’s own lookahead label in Table 4 (not a matched-label head-to-head), yield W =0, Holm-corrected pHolm=0.00586, r=0.886 on both cohorts. A physics- weight ablation isolates a cohort-asymmetric physics contribution. Survival evaluation yields a pooled Harrell C-index of 0.719 (Cox on the encoder embedding with principal component analysis (PCA)-32 fit on training windows only, scored on held-out test windows). The inverse probability of censoring weighting (IPCW)-weighted Brier score is 0.082; Schoenfeld p=0.210 on the held-out test partition. A preliminary cross-original equipment manufacturer (OEM) transfer probe (non- headline protocol) returns near-chance area under the curve (AUC) (0.514 [0.491, 0.537] for Physics-informed+Cox). Cross- OEM domain adaptation remains insufficiently addressed in the present benchmark. False-positive rate (FPR)-budgeted alarms (≤ 5/turbine/year) link the detector to wind-farm energy management: fewer false stops raise time-based availability and reduce curtailment from unnecessary technician visits.