Fifteen-Second Data Windows Could Transform Wind Turbine Lifetime Assessment

Every wind turbine tower is quietly accumulating damage. Each gust of wind, each pitch adjustment, each start-up and shutdown leaves behind microscopic fatigue cracks that, over years, determine whether a turbine can safely keep spinning beyond its design life. Deciding whether an ageing turbine deserves a life extension is one of the most consequential financial and safety decisions an operator can make, and it hinges on a single number: the accumulated fatigue damage in the steel. A new study published in Wind Energy Science shows that this number can be computed with remarkable accuracy using a tiny fraction of the data previously thought necessary, opening the door to fatigue monitoring across entire wind farm fleets rather than just a handful of instrumented machines.

The research, led by Catarina Miranda Oliveira of the University of Porto and the National Laboratory for Civil Engineering in Portugal, together with colleagues at the company Nadara, tackles a stubborn bottleneck in structural health monitoring. The gold standard for fatigue assessment is rainflow cycle counting applied to continuous, high-frequency strain measurements from sensors bonded directly to the tower wall. This method faithfully reconstructs the full spectrum of stress cycles the steel experiences, including large, slow cycles that span many minutes. But storing and processing 50 hertz strain data continuously, month after month, produces data volumes and computational burdens that are simply impractical to sustain across hundreds of turbines.

For decades, the industry has therefore relied on a compromise: the 10-minute statistical summaries that supervisory control and data acquisition, or SCADA, systems already collect fleet-wide. The problem is that 10-minute windows truncate stress cycles that stretch across window boundaries. These long-period cycles, driven by slow shifts in wind speed and direction, operational transitions, control actions, and daily or seasonal temperature swings, turn out to be disproportionately important. Because fatigue damage scales with a high power of stress amplitude in welded steel, a few large-amplitude slow cycles can outweigh thousands of small ones. Previous studies have found that ignoring them can bias fatigue estimates downward by anywhere from roughly 8 percent to more than 60 percent, a systematic non-conservatism that could mean a turbine is retired too early, or worse, kept in service on falsely optimistic numbers.

The Portuguese team had access to something rare: six months of continuous 50 hertz strain data from two instrumented sections of an onshore turbine tower, one near the base and one higher up, at a wind farm perched on a mountain ridge in central Portugal. The ridge-top site, with its irregular karstic terrain and quasi-linear turbine layout, produces heterogeneous, turbulent flow that makes it a demanding test case. The monitoring campaign, running continuously since September 2023, provided the raw material for a systematic benchmark: continuous rainflow counting as the reference baseline, conventional window-based counting, a residual-preserving framework known as low-frequency fatigue dynamics, and two radically compressed extrema-sequence representations.

The low-frequency fatigue dynamics approach is the conceptual heart of the study. Rather than discarding the unclosed half-cycles left dangling at each window boundary, the method preserves them as residual stress sequences, stitches the residuals together across windows, and runs rainflow counting on the concatenated result. The outcome is mathematically equivalent to processing the full continuous signal, but it can be computed window by window. The researchers’ modified implementation goes further by explicitly separating the damage contribution of cycles that close within a window from those reconstructed across boundaries, giving operators a granular picture of where fatigue is actually accumulating. Using this framework, the team recovered 99.5 percent of the baseline fatigue damage, regardless of how long the aggregation window was.

Perhaps the most counterintuitive finding concerns window length itself. One might assume that shrinking the conventional 10-minute window to a mere 15 seconds would automatically improve accuracy by capturing faster fluctuations. The data say otherwise. With 15-second windows and conventional counting, only 25.8 percent of the baseline damage survived, because shorter windows truncate even more of the slow, large-amplitude cycles that dominate accumulated damage. Shortening the window does not fix the truncation problem; it aggravates it. What matters is not the window length but whether the method preserves cycle continuity across boundaries. Once the residual-preserving framework is in place, the window length becomes almost irrelevant to total damage, though it dramatically reshapes the split between within-window and boundary-spanning contributions. At 15-second segmentation, fully 99.5 percent of identified damage migrated into the inter-window residuals, a redistribution caused by segmentation itself rather than any physical change in the loading.

The truly surprising result, however, comes from the extrema-sequence methods, which dispense with most of the signal entirely. Rainflow counting depends only on the sequence of stress turning points, not on the points in between. Exploiting this, the researchers tested two compressed representations. The start-peaks-valleys-end approach keeps, for each window, the initial value, all local peaks and valleys, and the final value, then concatenates these into a single sequence for one rainflow pass. The maximum-minimum approach is even more brutal, retaining only the highest and lowest stress values per window. Against the continuous baseline, the start-peaks-valleys-end method recovered 98.8 percent of the fatigue damage while storing roughly half the raw data, and the maximum-minimum method achieved 96.2 percent accuracy at 15-second windows while keeping just about 2 percent of the original data volume.

The compression numbers translate directly into deployment economics. Retaining only the concatenated residual sequence within the low-frequency fatigue dynamics framework requires approximately 7 percent of the raw data volume, yet suffices to reconstruct reference-equivalent damage. For an operator weighing whether to instrument an entire fleet, the difference between storing continuous 50 hertz strain data and storing a handful of extrema per 15-second interval is the difference between an impractical research exercise and a routine operational pipeline. The study also quantified the cost of downsampling: reducing the sampling rate from 50 to 25 or 12.5 hertz introduced negligible deviation, while a deliberately conservative 6.25 hertz scenario underestimated total damage by up to about 9 percent, with the side-to-side direction, governed by faster rotor-rotational content, proving more sensitive than the fore-aft direction dominated by aerodynamic thrust.

The choice of fatigue standard emerged as another consequential factor. Comparing Eurocode 3 detail category 80 against the DNVGL-RP-C203 welded-steel curves, both bi-linear formulations with slopes of 3 and 5, the team found that Eurocode 3 yielded damage estimates 17 to 66 percent lower than DNV depending on tower section, with the gap widening higher up the tower where stress-range distributions are dominated by lower-amplitude cycles. For the welded tubular details typical of onshore towers, the authors identify Eurocode 3 as the more appropriate standard, but the discrepancy underscores how much inferred remaining life can depend on a seemingly bureaucratic modelling choice rather than on the physics of the structure itself.

The implications reach well beyond a single turbine on a Portuguese ridge. Because second-resolution SCADA metrics are now deployable at scale and reasonable cost, the study establishes a practical pathway to SCADA-aligned fatigue monitoring at 15-second resolution across whole fleets, without continuous high-frequency data storage or processing. The authors point toward predictive extensions: forecasting extrema sequences, particularly the highly compressed maximum-minimum representation, or predicting the low-frequency residual sequence directly, which at short windows carries the large majority of fatigue damage while representing only about 7 percent of the raw data. Open questions remain, including extrapolating damage estimates from instrumented sections to the tower base or top, decomposing fatigue by operational state such as start-up and idling transients, and assessing wake effects across turbine arrays. But the core message is clear: the fatigue information that decides a turbine’s fate survives aggressive compression, and that insight could reshape how the ageing wind fleet is managed, one 15-second window at a time.

Subject of Research: Sub-minute SCADA-aligned fatigue monitoring methods for wind turbine tower lifetime assessment

Article Title: Sub-minute fatigue monitoring for enhanced lifetime assessment of wind turbines

Article References: Oliveira, C. M., Biscaya, A., Santos, J. P., Caetano, E., & Xu, M. (2026). Sub-minute fatigue monitoring for enhanced lifetime assessment of wind turbines. Wind Energy Science, 11(10), 3841-3862. https://doi.org/10.5194/wes-11-3841-2026

Image Credits: AI Generated

DOI: 10.5194/wes-11-3841-2026

Keywords: wind turbines, fatigue monitoring, structural health monitoring, rainflow cycle counting, SCADA, lifetime extension, strain measurements, low-frequency fatigue dynamics, tower fatigue, wind energy, Eurocode 3, DNV