This article also appears in
Subscribe now »

The 12-m open-return subsonic wind tunnel consists of a centrifugal fan, a diffusing section, a settling chamber, a contraction section, and a working section.

ETS researchers develop new methodology for wind tunnel calibration

In recent years, the École de Technologie Supérieure (ETS) Research Laboratory in Active Controls, Avionics, and Aeroservoelasticity (LARCASE) has acquired several research apparatus—a research flight simulator Cessna Citation X, an open return subsonic wind tunnel, and a flight autonomous system (UAV)—making it one of the few multidisciplinary research laboratories in Canada with a wide range of equipment with the capabilities of simulating aircraft models, especially airfoils, and validating the models with experimental data collected on the ground (wind tunnel) and in-flight (UAV).

LARCASE researchers have developed a new approach for wind tunnel calibrations using a limited number of dynamic pressure measurements and a predictive technique based on Artificial Neural Network (ANN). Complementing the development of an ANN model, a further objective of the research was to investigate an optimal method for determining local flow characteristics by means of X, Y, and Z coordinates. Using a minimum of data collected from the wind tunnel calibration for training, the ANN model would generate the proper pressure for any given 3-D coordinate inside the test section.

The 12-m open-return subsonic wind tunnel is a research apparatus used to test airfoils and validate CFD models. The wind tunnel allows for a safe control of the flow conditions and makes measurements of pressure distribution on a wing shape possible. At the beginning, the air pressure rises through the centrifugal fan creating lateral mixing of fluid layers, then the turbulent particles are straightened by the filters producing a laminar flow. The wind tunnel consists of a centrifugal fan, a diffusing section, a settling chamber, a contraction section, and a working section.

The flow develops a 0.18 Mach maximum speed due to the engine and the double impeller centrifugal fan. The two inlets at the opposite side of the rotor allow the air supply to increase the pressure flow, and the use of the 24 small propellers by impeller allows the fan to turn at a much higher speed than normal fans with large blades. The engine and the centrifugal fan are located inside the soundproof mechanical room protected from debris and dust.

The diffusing section consists of a wide-angle diffuser, a large settling chamber, a contraction section, and a test section. From the static pressure buildup the flow is projected to an oval shaped circular pattern flow straightener. Then the flow goes through a series of five filters, the first is a honeycomb-shaped filter, and the other four are nylon squared shape filters positioned 0.5 m from each other.

The settling section allows the flow to go from a turbulent state to a laminar flow.

LARCASE's wind tunnel has two test sections, the main one is 0.6 x 0.9 m made in wood with Plexiglass removable doors able to reach 0.12 Mach. The second test section, half the volume of the first one, is able to reach 0.18 Mach.

LARCASE researchers used the Extended Great Deluge (EGD) algorithm in hybridization with the neural networks to find the predicted pressure inside of the test chamber of the open return subsonic wind tunnel.

The hybrid NN-EGD method is proposed to control the pressure distribution, by varying the coordinates of the point inside the test chamber, wing speed, and temperature. The EGD algorithm is used to obtain the optimal network configuration such that the error is as small as possible. Qualitative performance measures are used that describe the learning abilities of a given trained neural networks.

To train and test the NN, ANSYS Fluent software is used to determine the pressure values inside the test chamber. The coordinate (X, Y, Z), the wind velocity (V), and the temperature (T) of the test chamber represent the inputs of the model, the output is the pressure. A total of 81628 points are used to train, validate, and test NN-EGD. The validation data represent 15 % of the data set, 15% of the data set to test the approach, and the rest to train NN. These points are selected randomly. Using the EGD algorithm, many architectures are tested. The objective was to obtain the simplest configuration to give the best results in a short time of compilation. After randomly trying different combinations of numbers of neutrons and layers, the best results are obtained using a NN architecture composed of four layers feed-forward network, The number of neurons in each layer is 12, 15, 10, and 1, respectively. The NN inputs are X, Y, Z, V, and T. The output is the pressure.

The optimal architecture obtained using the EGD algorithm and obtaining the best results is composed of four layers feed-forward network. The NN-EGD are implemented in Matlab.

To test the approach, researchers used 14440 points. The average error of the obtained results in plane 1 is equal to 7.22 %. In the plane 2, the error is 4.42%, and the error in the plane 10 is equal to 3.16 % of the theoretic pres-sure.

By using this approach, the researchers successfully obtained the value of the pressure in each point of the dataset according to the coordinates of each point, the wind speed, and temperature of the test chamber.

Continue reading »