From Few Instrumented Turbines to the Whole Farm: Fatigue Estimation Using Learned Acceleration Representations and LSTM Regression
Yacine Bel-Hadj, Francisco de Nolasco Santos, W. Weijtjens, C. Devriendt
Strain gauges provide the most direct route to fatigue estimation because they allow computing the Damage Equivalent Moment through cycle counting. Their cost and the need for regular maintenance make them expensive to operate in offshore environments. As a result, strain gauges are installed on only a few fleet leaders, leaving most turbines without ground truth on accumulated fatigue. Attempts to compensate for this lack of strain data have relied on learning a mapping from 10 minute SCADA aggregates to DEM and extrapolating it to the rest of the farm, but SCADA channels are often incomplete, and their low temporal resolution removes the fast loading cycles that dominate fatigue. Proxies such as the standard deviation of power are used to approximate fluctuations in mechanical loading, but the information they provide is coarse. High frequency acceleration preserves the full vibration response and captures mechanical loading fluctuations that 10 minute SCADA cannot resolve, making it a viable surrogate when strain is absent. The objective of this contribution is then to learn a mapping from accessible high frequency acceleration to DEM and to extrapolate it from the few SG-instrumented turbines to the full farm. The method has two stages. First, an autoencoder is pretrained in an unsupervised manner on triaxial acceleration from forty-four turbines of a single offshore wind farm. Each 10-minute record is split into shorter windows, so the encoder learns local dynamic patterns at patch level. Then, domain adversarial neural network regularization suppresses turbine specific signatures and aligns the latent space across assets, producing features that remain stable across the fleet. After pretraining, the encoder is frozen. Second, for each 10-minute interval, the windows are embedded with the frozen encoder to form a sequence of latent vectors. This sequence is passed to an LSTM trained only on the strain instrumented turbines, and the final hidden state for each 10-minutes of acceleration is fed to a small neural network for regression. One instrumented turbine is held out to test transferability. The trained Encoder-LSTM reaches an R2 of about 0.97 on the held-out turbine, outperforming SCADA based and hand-crafted acceleration feature baselines that remain near 0.95. The latent space aligns with wind speed, pitch angle and other operational variables, confirming that the encoder extracts the relevant physical behaviour. The performance gain comes from the higher temporal resolution of the vibration data, which preserves dynamic load cycles that SCADA aggregates remove. All results presented here are obtained from accelerometer data alone, without any need for SCADA, metocean, or additional measurements. This gives a single source workflow that avoids data synchronization and missing channel issues. A small set of strain-equipped turbines provides the ground truth needed to train the mapping, and high frequency acceleration is increasingly available as the default sensing setup across modern wind farms. The result is a farm-wide fatigue estimation framework that converts sparse strain information into full coverage and moves population-based Structural Health Monitoring closer to practical deployment.