| Popis: |
Virtual sensing allows to estimate quantities that are too costly or difficult to directly measure, by combining measurements from more practical sensors and locations with models of the system in question. Here, the application is the estimation of the load (forces and torques) exerted by a wind turbine blade on the pitch bearing that connects it to the rotor hub. This article focuses on selecting the sensors (types and positions) to provide the measurements to the estimator. Three types of sensors are considered: accelerometers, displacement sensors, and strain gauges, from cheapest and noisiest to most expensive and accurate. These must be placed at accessible locations on the assembly surrounding the pitch bearing to obtain the best estimation of the load at the lowest price. Sensor selection is crucial to the performance and cost of a virtual sensing solution. Typically, the design and performance of the estimator depends on the choice of sensors while the sensors are chosen based on estimator performance. Because of this coupling, an initial suggestion of sensor set made independently of any existing estimator is useful. This work uses a reduced order model of the pitch bearing assembly, based on an analytical bearing model and a finite element model of the surrounding assembly. It is observed that this model is sufficiently linear for the considered loads that a lower-bound on the estimator uncertainty can be derived using Riccati’s equation for the asymptotic state covariance of a Kalman Filter. This provides a metric for the performance of a given sensor set, based only on the model of the pitch bearing assembly and the positions and accuracies of the individual sensors. This single metric is valid in all operating conditions where the model remains accurate, and does not require simulation results from one or more representative scenarios. This metric is used in combination with sensor cost to form the objective function of the sensor selection problem. This sensor selection problem is combinatorial, with discrete decision variables and a black-box objective function. For this reason, black-box optimization techniques are used to explore a large number of possible sensor sets that are then combined and pruned to obtain suggestions of good sensor sets. Optimality cannot be guaranteed for this type of problem but the presented approach outperforms hand-picked sensor sets using typical heuristics (equally distributed positions around the bearing assembly, positions with highest deformations, anti-nodes of resonant modes, …), while remaining consistent with zones of higher sensitivity around the pitch bearing assembly. Some human supervision is still required because the objective function does not consider some more practical considerations in terms of sensor placement and to tune the sensor selection towards acceptable sensor sets. ispartof: pages:4594-4607 ispartof: Proceedings of ISMA2022 including USD2022 pages:4594-4607 ispartof: 2022 Leuven Conference on Noise and Vibration Engineering (ISMA2022) location:Leuven, Belgium date:12 Sep - 14 Oct 2022 status: published |
