Predictive Maintenance · Wind Energy

Detect drivetrain faults
days before failure.

TorqueScope identifies emerging faults in wind turbines up to 59 hours before breakdown - using only the SCADA data you already collect. No new sensors. No historical failure data. No ML team required.

View Evidence Read the Method

Method

Physics-grounded detection,
not black-box ML.

Two independent detectors, each grounded in physical principles, must agree before an alarm is raised. Neither requires fault examples. Neither requires domain expertise to deploy.

Step 01

Lomb–Scargle Periodogram

A 7-day sliding window over temperature sensors is decomposed into its frequency components. Healthy drivetrain signals show stable periodic structure - rotor harmonics, diurnal thermal cycles, mechanical resonances. Deviation from this baseline triggers the heuristic score.

score ← f(amplitude_ratio, residual_ratio, CV)

Step 02

Normal Behaviour Model

Expected temperature is modelled as a function of power output and ambient conditions across a 200-bin operational grid (20 × 10 bins). Residuals from this physics-informed lookup reveal anomalous thermal behaviour invisible in raw sensor streams.

residual ← T_actual − T_expected(P, T_amb)

Step 03

Hybrid Criticality Engine

When both detectors agree, confidence is amplified. A criticality counter accumulates evidence over time, rising on anomaly readings and decaying otherwise. Single-detector signals are suppressed by 50%, converting noise into structured forewarning.

alarm := criticality ≥ 72

Evidence

Four faults. Three farms.
All detected early.

Validated against the CARE benchmark - 95 datasets, 3 anonymised wind farms, ground-truth fault timestamps confirmed by maintenance personnel. Lead time measured from first TorqueScope alarm to documented fault onset.

#	Fault Description	Farm	Sensors	Lead Time
01	Gearbox failureDrivetrain · bearing degradation	Farm A - Onshore, Portugal	86	32.3hours
02	Yaw grease pump failureYaw system · lubrication loss	Farm C - Offshore, Germany	957	50.0hours
03	Gear oil pump coupling defectLubrication · coupling wear	Farm C - Offshore, Germany	957	59.3hours
04	Main bearing damageBearing · rolling element fatigue	Farm B - Offshore, Germany	257	29.5hours

NOTE

Lead times reference ground-truth maintenance records, not model confidence thresholds. "Fault confirmed" denotes the timestamp at which wind farm personnel documented the physical failure - not when the algorithm assigned a probability.

Signal

Hybrid anomaly score -
Scenario 01 (Gearbox failure)

Hybrid score across the prediction window for Farm A, Event 10. The criticality counter crosses threshold 72 a full 32.3 hours before the documented fault onset.

Hybrid anomaly score · Farm A · Asset 10 · CARE Event 10

Score

Threshold 0.475

First alert

Fault confirmed

Advantage

Zero hardware.
Zero historical failures.
Operational from day one.

CMS installations require hardware procurement, site visits, and years of failure data to train on. TorqueScope connects to your existing historian in days.

→

No new sensors required

CMS hardware costs €5,000–15,000 per turbine and requires a site visit for installation and calibration. TorqueScope works from any SCADA historian export.
→

Zero-shot from day one

The Normal Behaviour Model calibrates on healthy baseline operation alone. There is no requirement to have ever observed a fault - making it viable for new farms, new markets, and operators of any size.
→

Earliness, not just detection rate

Industry benchmarks reward detection rate. TorqueScope is designed to maximise how early. A fault caught 59 hours offshore is a planned maintenance window - not an emergency €200K repair and multi-day downtime event.
→

Interpretable physics, not opaque models

Every alarm traces back to specific sensors and physical quantities. No fleet-wide training, no cross-operator data sharing. Interpretable outputs compatible with existing O&M workflows.

Capability	Typical CMS	TorqueScope
New sensors needed	Yes - €5–15K/turbine	None
Historical fault data	Required	Not required
Deployment timeline	Months	Days
Viable on small farms	Uneconomic	Yes
Alarm interpretability	Black-box	Sensor-level
Max validated lead time	Varies	59.3 hours

CARE

Benchmark score: 0.588 on the CARE metric across 95 datasets. Autoencoder baseline: 0.66. The 11% gap reflects missed weak-signature faults - an expected and documented limitation of a zero-training-data approach.

Detect drivetrain faultsdays before failure.

Physics-grounded detection,not black-box ML.

Four faults. Three farms.All detected early.

Hybrid anomaly score -Scenario 01 (Gearbox failure)

Zero hardware.Zero historical failures.Operational from day one.

Connect your farm.Get your first alert.

Detect drivetrain faults
days before failure.

Physics-grounded detection,
not black-box ML.

Four faults. Three farms.
All detected early.

Hybrid anomaly score -
Scenario 01 (Gearbox failure)

Zero hardware.
Zero historical failures.
Operational from day one.

Connect your farm.
Get your first alert.