UC Berkeley

The structural health of wind turbine towers is not well understood and requires manual maintenance that involves visual inspection, bolt tightening, and interruptions to power generation. This naturally leads to the question of how to better understand the turbine tower's health without manual intervention. In this work, we propose a solution from sensing to inference. Distributed fiber optic sensing (DFOS) is an emerging technology for measuring mechanical strain. Unlike traditional point-wise sensing, DFOS enables spatially continuous measurements along a single sensor, resulting in a more comprehensive dataset. The two technologies examined are optical frequency-domain reflectometry (OFDR) and phase-sensitive optical time-domain reflectometry (ϕ-OTDR), which is a technology used in distributed acoustic sensing (DAS). OFDR is a tested and proven strain measurement technology commonly used for structural health monitoring, but can only measure strain over short distances (10s of meters). OFDR is first used to validate measurements obtained with ϕ-OTDR, which can measure at much longer ranges (several kilometers). Due to its sensing range, ϕ-OTDR is a promising technology for monitoring a network of many wind turbines via a single fiber-optic cable. Before real-world deployment, these technologies, as well as their ability to detect structural phenomena associated with loose bolts and material damage within the tower, need to first be evaluated and validated on real mechanical systems. To do so, we perform experiments on a model wind turbine and then a full-scale wind turbine on a shake table. The results show good agreement between ϕ-OTDR and OFDR measurements and show the respective strengths of the technologies’ abilities to capture local and global structural phenomena. After validation, an onshore wind turbine was then instrumented, and strain measurements were collected over the span of a year. Classical system identification methods were applied to assess the structure's health. However, due to limited input data, the usable algorithms were restricted to output-only methods. To address this issue, we propose a differentiable structures framework that takes a step towards solving physics-constrained inverse problems via differentiable programming, addressing the deterministic point-estimate aspect of the input-parameter-state estimation problem and leveraging advances in system identification and automatic differentiation. We pose the estimation problem as a partial differential equation constrained optimization problem, employing gradient-based optimization and automatic differentiation to compute the necessary gradients. We recontextualize finite element model updating within partial differential equation constrained optimization, drawing insights from the theory built in the optimization community. To evaluate our proposed framework and methodology, we perform computational experiments on a finite element model of a wind turbine subjected to stochastic dynamic inputs. The framework yields accurate pointwise estimates of inputs, parameters, and states on both synthetic and real datasets generated through simulation and from the instrumented onshore wind turbine, respectively. By leveraging system identification methods and experiments especially designed to excite certain modes of the structure, the estimation problem is decoupled, thereby allowing for faster optimization. The differentiable forward model allows for subsequent uncertainty quantification, which is enabled by this design. We additionally explore utilizing machine learning concepts, namely neural reparameterization, on a real dataset to help improve optimization in the inverse input estimation setting. Critically, this method enables efficient gradient-based optimization for high-dimensional estimation problems.