Transfer Learning Approach to Predicting Operating Conditions During Wind Tunnel Tests

Assuring the quality and reliability of data collection in wind tunnel tests requires accurately determining the operating conditions, such as onflow speed and angle of attack. Such operational data are conventionally attained using measurement systems including pressure and optical sensors, and the accuracy predominantly relies on proper calibration and considering uncertainties. To mitigate the impacts of these factors while enhancing system reliability via redundancies, computational methods such as deep learning techniques which resulted in major breakthroughs in many research areas can alternatively be leveraged. For this purpose, we specifically consider convolutional neural networks (ConvNets) to learn the mapping from surface pressure data to onflow speed and angle of attack. Furthermore, with the goal of first bringing the knowledge learned from computational fluid dynamics (CFD) data to the wind tunnel testing phase with streaming data, a transfer learning approach is proposed and demonstrated for a wind turbine blade tip. This approach requires first training a neural network based architecture offline by using only CFD data, then freezing the weights of several layers, and retraining the remaining layers using only wind tunnel data. Using this approach, the feasibility of knowledge transfer from CFD runs to wind tunnel tests as well as real-time prediction and online learning during the experiments are successfully demonstrated.