Faithfulness-audited LLM alerting for multi-scale SCADA-based anomaly detection and subsystem localization in wind turbines

Supervisory control and data acquisition (SCADA) data enable wind-turbine condition monitoring without additional sensors. However, existing stationarity-based detectors often evaluate each signal over a single fixed window, monitor parameters in isolation, and issue only binary alarms, limiting their deployment value. This paper presents a deployment-oriented energy-informatics framework that extends the augmented Dickey–Fuller stationarity paradigm across temporal scales, physical channels, and alert granularity. A causal multi-scale engine evaluates each signal over several sliding windows, with control limits frozen on a healthy commissioning period. Evidence is then fused across physically related channels to localize anomalies to subsystems from residual stationarity rather than magnitude, and a deterministic cascade escalates alerts through suspicion, confirmed, and reinforced states that require cross-channel corroboration. A constrained language model verbalizes the structured evidence into tiered, operator-facing warnings; it performs no detection or diagnosis, while a deterministic faithfulness guard blocks unsupported subsystem, severity, or fault-mode claims. Under a causal protocol without look-ahead, the framework detects a gearbox failure 63 h before its labelled onset and, on a year-long dataset, flags the fault inside the labelled failure window and 146 h before that window ends, localizing both to the drivetrain, with no confirmed false alarm over the monitored-healthy period. Across 880 audited generations from four language models, no warning fabricated content beyond the engine evidence.