There is a growing need for effective fatigue estimation in offshore wind turbines, as it is crucial for extending farm life and ensuring that newer turbines meet their design lifetimes. Previous methods have relied on SCADA (Supervisory Control and Data Acquisition) data, which provides valuable insights into environmental and operational conditions (EOC). However, access to high-quality SCADA might not always be a given. Without SCADA, predicting fatigue becomes challenging, as critical operational parameters such as wind speed and rotor speed are unavailable. To overcome this, this study explores the potential of using only dedicated structural monitoring accelerations for fatigue estimation on the foundations of offshore wind turbines. It proposes a novel approach leveraging autoencoders (AEs) to analyze acceleration spectra in an unsupervised manner through representation learning. AEs embed acceleration data into a latent space, capturing fatigue-sensitive features while also providing operational insights into the turbine’s state. The research evaluates the ability of the latent space to infer EOC and serve as a surrogate of SCADA data for fatigue estimation. The findings demonstrate that AEs can successfully identify fatigue-relevant patterns from acceleration data alone, enabling accurate fatigue predictions in real-world scenarios and significantly improving on baseline methods. This approach offers a promising alternative for structural health monitoring when SCADA data is unavailable.