Semi-Supervised Learning for Geotechnical Soil Property Estimation in Offshore Windfarm Sites

Abstract Site characterization and monitoring of the subsurface formations around wind turbine locations are crucial for reliable wind farm construction, operation and maintenance. In order to extract relevant information about subsurface soils, ultrahigh-resolution (UHR) seismic survey and geotechnical cone- penetration testing (CPT) is often acquired, processed, interpreted and integrated, which could be repeated over time for site monitoring purposes. Due to the size of the area to be investigated and the manual efforts to complete multiple steps in the traditional workflow, the turnaround time for soil property estimation in a wind farm site can be quite long. In this study we implement a semi-supervised learning workflow to automate the task, which integrates URH seismic and CPT logs through two convolutional neural networks (CNNs), with one for seismic denoising and feature engineering (SDFE) and the other for seismic-CPT integration (SCI), which reduces the difficulties in CNN training due to poor data quality and small data quantity. The two components are connected by implementing the encoder of the pretrained SDFE-CNN as part of the SCI-CNN encoder. As tested on a public wind farm site, the use of deep learning leads to promising results in terms of both quality and efficiency. The proposed workflow is also extensible to include additional information, such as structure and velocity models, for further constraining the SCI-CNN. Highlights: A semi-supervised learning workflow is proposed for soil property estimation from UHR seismic and CPT tests in a wind farm site,allows estimating the essential soil properties such as cone-tip resistance from post-stack UHR seismic as tested on a real windfarm site HKZ, andreduces the turnaround time of windfarm site characterization compared to traditional workflows.