Start with the failure mode, not the algorithm
A drivetrain failure on a modern turbine is rarely a surprise to the machine, only to the operator. Gearbox bearings, main bearings, and generators degrade over weeks to months, and they shed evidence the whole way down: heat the thermodynamics can't explain, vibration at characteristic defect frequencies, metal particles in the oil. The engineering question is not whether the signal exists. It is whether your data, your models, and your maintenance organisation are set up to catch it and act on it.
That starts with being specific about what you are predicting. Three component groups dominate major-correctives on most fleets: the gearbox (particularly high-speed-shaft bearings and the planetary stage), the main bearing, and the generator. Each has a different degradation signature and a different realistic detection horizon. Main bearings turn slowly and damage accumulates gradually, so condition monitoring practice routinely reports lead times of many months. Gearbox bearing faults typically give weeks to a few months of warning; generator bearing problems can progress faster. "Weeks in advance" is not a property of the model. It is a property of the failure mode, and it determines which data you need before you write a line of code.
The data hierarchy
There are three tiers of data, and they are not interchangeable.
Ten-minute SCADA aggregates are the universal baseline: temperatures, power, wind speed, rotor speed, pitch angle, averaged over ten-minute windows. They cost nothing extra and every turbine has them. Temperature channels are the workhorse. A gearbox bearing running warmer than its operating conditions justify is one of the most robust early indicators in the field, and published case studies regularly detect such anomalies one to three months before failure. The limitation is physics: ten-minute averaging destroys the high-frequency content that mechanical faults actually live in, so SCADA tells you something is wrong long before it tells you what.
High-frequency vibration from a condition monitoring system (CMS) fills that gap. Accelerometers on the drivetrain sampling in the kilohertz range, processed with spectral and envelope analysis, reveal energy at bearing defect frequencies (inner race, outer race, rolling element) and can localise the fault to a specific bearing at an early stage. This is the earliest and most diagnostic warning available.
Oil chemistry and debris data confirm what vibration suspects. Inline particle counters trending ferrous debris tell you wear is active, not historical, and the rate of particle generation is one of the better inputs for staging severity.
The emerging-market reality is that many fleets have only the first tier. Retrofit CMS carries meaningful per-turbine capex, and on older or smaller machines the case is genuinely marginal. The pragmatic architecture is a fleet-wide SCADA-based screen on every turbine, with CMS retrofits prioritised on the machines where a drivetrain failure hurts most: the largest, the oldest, the hardest to reach with a crane. On latency: for weeks-ahead detection, a daily batch pipeline is entirely adequate. Do not build streaming infrastructure to answer a question that changes once a day.
Modelling: residuals before deep learning
The workhorse technique is the normal-behaviour model: train a regression model, gradient-boosted trees or a small neural network, to predict a target signal such as gearbox bearing temperature from operating conditions (power, ambient temperature, rotor speed, nacelle temperature), using only periods of known-healthy operation. In service, monitor the residual between predicted and actual. A healthy turbine's residual hovers around zero; a degrading component drifts. Sustained drift, tracked with something as simple as an exponentially weighted moving average with control limits, is the alarm, not any single excursion.
Autoencoders generalise the same idea: reconstruct the full sensor vector and use reconstruction error as a multivariate health index. They help when you do not know in advance which channel will move, at the cost of interpretability. And interpretability is what gets a technician into a truck.
Supervised classification, typically gradient boosting on engineered features like inter-stage temperature deltas, vibration band energies, and particle-count rates of change, works well in papers and struggles in fleets, for a simple reason: a hundred-turbine fleet may produce only a handful of relevant failures per year. That is not a training set; it is a validation set. Remaining-useful-life estimation sits at the top of the ambition curve, and honesty is warranted: RUL intervals on real fleets are wide. In practice, operations teams extract more value from severity staging (watch, inspect, plan replacement) than from a point estimate of days remaining.
Labels are the bottleneck
Every supervised result rests on labels, and in this domain the labels are maintenance work orders, which are noisy in every way that matters. They are free text in inconsistent formats, component coding varies by technician, and the recorded date reflects when the repair happened, not when the fault initiated. Deciding that a failure "began" 60 days before the replacement work order is a modelling assumption dressed up as ground truth, and it directly shapes what the model learns. Add severe class imbalance, with years of normal operation per failure event, and it becomes clear why label engineering usually consumes more effort than model training. The practical response is twofold: treat work-order curation as a first-class engineering task with per-failure-mode pre-failure windows, and lean on unsupervised and normal-behaviour approaches so that scarce labels are spent on validating the system rather than training it.
Lead time, false alarms, and the price of crying wolf
Sensitivity buys lead time and false alarms together; there is no setting that gives you one without the other. The economics are asymmetric. A missed gearbox failure can mean a crane mobilisation and months of lost production, while a false alarm costs a borescope inspection and an oil sample. But false alarms compound. A model that pages the operations team weekly with nothing behind it is dead within a quarter, whatever its recall. Trust is built on precision.
And an alert only creates value if it changes a maintenance decision. That means every alert should carry:
the suspected component and failure mode, not just an anomaly score
the evidence: residual trends, spectra, particle counts a reviewer can inspect
a severity stage mapped to a defined action (monitor, inspect, or plan intervention)
the track record of similar alerts on the fleet
It also means operational readiness: borescope capacity, a spares strategy, and crane logistics arranged in advance. In many emerging markets, crane mobilisation lead times run to months, which is precisely why weeks of warning matter. The prize is converting an unplanned catastrophic exchange into a planned up-tower repair executed on your schedule.
Where this goes next
The tooling will keep improving: cheaper retrofit sensors, better exploitation of high-frequency SCADA, fleet-level models that transfer learning across turbines and sites. None of it changes the discipline. Match data to failure mode, prefer interpretable residuals, treat labels as engineering, and set thresholds from maintenance economics rather than ROC curves. A prediction is only intelligence. The value is the crane you never had to mobilise on emergency terms, and that is the standard every predictive maintenance programme should be measured against.
