A wind farm operator runs 3,200 turbines across 14 sites. Each turbine streams 180 sensor channels: vibration, temperature, oil pressure, blade pitch, generator RPM: at 1Hz. An unplanned gearbox failure costs roughly $250,000 in parts, crane hire and lost generation. A planned replacement during a scheduled maintenance window costs about $60,000.
Over four years of history there have been 71 gearbox failures across the fleet. Maintenance logs are inconsistent: some record the failure date, some the repair date, and some only "gearbox serviced".
The operator wants 30 days of warning so a replacement can be folded into a scheduled window. They currently run fixed-interval maintenance and are replacing many healthy components.
Sites have intermittent satellite connectivity: some go offline for hours.
Build the architecture on a canvas: place the components, configure them, connect them into a data flow, and write a short reason for each one. The AI reviewer grades your design against a rubric written specifically for this problem.
You have 71 failures and 3,200 turbines streaming 180 channels. Why is a deep model on raw sensors the wrong instinct here?
Maintenance logs disagree with each other about when a failure happened. How do you build labels you trust?
Your model alerts 3 days before failure. Is that a success? Why does the answer change the metric you report?
Minimum 6 components · needs a wide desktop screen