Researchers built an AI system that talks to wind turbine operators, and predicts failures before they occur
Wind turbines are among the most mechanically demanding machines in modern energy infrastructure — massive rotating structures exposed around the clock to temperature swings, vibration, and unpredictable loads. When something goes wrong, the costs compound quickly: halted production, emergency crews dispatched to remote sites, and components that can take days to replace.
Existing monitoring systems can log what’s happening inside a turbine. What they can’t do is interpret it — or warn an operator that a gearbox is trending toward failure before it actually fails.
A new system developed by researchers combines real-time IoT sensors, machine learning, and a conversational AI interface that doesn’t just flag anomalies — it explains them.
An XGBoost model was first trained on a large industrial SCADA dataset covering 89 cumulative years of operational data from 36 turbines across three wind farms.
The problem with watching a turbine fail
Wind turbines don’t fail quietly. Gearboxes grind through years of torque cycles. Electrical systems absorb heat, vibration, and moisture. When something finally gives, the damage is rarely isolated — and the repair bill reflects that.
The standard response has been SCADA monitoring: systems that track operational variables and trigger alerts when readings cross fixed thresholds. The problem is that thresholds are static. They can’t interpret what it means when nacelle temperature climbs while power output drops and wind speed stays constant. Each sensor speaks alone; no one connects the conversation.
Most AI research into turbine health has tried to fill that gap, but with a narrower lens. Studies typically rely on a single data source and present findings through dashboards that operators read passively — one-directional, the machine never explaining itself.
A physical prototype wired for intelligence
Researchers at Dokuz Eylül University built a small-scale vertical axis wind turbine prototype instrumented with sensors well beyond what standard SCADA systems track: temperature, humidity, vibration, wind speed, power output, and both gas and acoustic sensors.
That last pair was deliberate. Gas sensors can detect early signs of electrical fire risk. Acoustic sensors pick up high-frequency micro-cracking before vibrations become structurally significant — failure signals invisible to conventional monitoring.
Two ESP32 microcontrollers handle data collection and transmission, pushing readings to the ThingSpeak cloud platform every 15 seconds, with analysis completing within 5 seconds. Data moves through a preprocessing pipeline that filters outliers, applies temporal averaging, and uses K-means undersampling to balance the heavily skewed ratio of normal versus failure events.
Teaching machines to spot failure, then teaching them to explain it
The modeling approach is a two-stage teacher–student transfer learning framework. An XGBoost model was first trained on a large industrial SCADA dataset covering 89 cumulative years of operational data from 36 turbines across three wind farms. That model became the “teacher” — labeling the prototype’s unlabeled sensor stream using only the variables the two datasets share.
A LightGBM “student” model was then retrained on the full prototype dataset, including gas, acoustic, and vibration inputs the teacher never saw. It reached 98.37% accuracy and a 99.16% F1 score. Feature importance analysis confirmed that gas and sound sensors were the strongest new predictors of failure.
For real-time inference, XGBoost was selected as the operational engine. Using seven key features identified through SHAP analysis — nacelle temperature, gearbox temperature, wind speed, and voltage among them — it achieved approximately 89.68% accuracy and a 95.65% ROC-AUC on industrial SCADA data. The researchers also introduced a metric called Overall Performance, designed to penalize models that hide high false alarm rates or missed detections.
When the turbine speaks plain English
The system’s most distinctive feature isn’t its accuracy — it’s what happens after a prediction.
An LLM interface powered by Google’s Gemini API is embedded directly into the Oracle APEX dashboard, letting operators ask questions in plain language and receive specific, contextualized responses about turbine health, failure risk, and recommended actions. The LLM doesn’t guess. Before any sensor data reaches the model, a deterministic safety matrix in the Python back end evaluates readings against known fault patterns. When a threshold is breached, pre-validated engineering context gets injected into the LLM’s prompt — the model synthesizes that into language rather than generating its own diagnosis from scratch.
In a simulated generator jamming test, the system matched high wind speed alongside near-zero voltage to its fault matrix, and the LLM issued a directive: halt the system and inspect blades and rotor-generator coupling. In a thermal overheating simulation, a nacelle temperature breach above 50°C triggered a step-by-step emergency shutdown protocol within seconds.
Promising results, honest caveats
The headline accuracy figures deserve careful reading. The prototype’s labels were generated by the SCADA-trained teacher model — not by physically verified destructive testing. That means the student model’s high scores partly reflect how well it learned the teacher’s decision boundaries, not independently confirmed ground truth. The researchers are explicit about this: the results are a proof-of-concept, not a certified industrial benchmark.
Scale is another honest constraint. The gap between a lab prototype and a full industrial turbine is significant — vibration signatures, thermal gradients, and electrical behavior may not transfer cleanly to real wind farm conditions. Consumer-grade sensors and ESP32 microcontrollers offer research value but may not survive years of outdoor deployment.
Future development points toward edge computing for offline inference, deep learning sub-models for classifying specific fault subtypes, and digital twin integration for 3D visual monitoring. Each step would bring the system closer to conditions where it’d need to perform reliably — not in a lab, but at the base of a turbine on a remote hillside, when something starts to go wrong.
The full study can be checked here: Saygı, İ. G., Mahyanbakhshayesh, S., Ak, B., Özcanhan, M. H., and Ünlütürk, M. S.: A real-time IoT, LLM, AI-supported wind turbine failure prediction system, Wind Energ. Sci., 11, 3983–4003, https://doi.org/10.5194/wes-11-3983-2026, 2026.
Daniel García is an Editor-in-Chief with strong expertise in structural work and engineering principles. He combines this technical foundation with deep knowledge of energy, spatial design, and emerging technologies, bringing a forward-thinking and analytical approach to editorial leadership.
