PSI - Issue 80

Tafara E. Makuni et al. / Procedia Structural Integrity 80 (2026) 105–116 Tafara E. Makuni / Structural Integrity Procedia 00 (2019) 000 – 000

113 9

methods. For each of these methods, the error was calculated as the difference between the actual testing data ‘output’ and the output given by the trained model. The results in Figure 7 show that the BRBP is the most accurate and computed the fastest compared to the LMBP and SCGBP methods.

Figure 6: ML model for aerodynamic load reconstruction conducted using the LMBP method to train the ANN.

Figure 7: Error analysis of ML model trained using the (i) BRBP, (ii) LMBP, and (iii), SCGBP.

This ML model was extended further to take as inputs; M ∞ , α ∞ , h and the pressure from 4 different locations on the upper surface, which represent the physical data from sensors on the wing section. For the aerofoil section, this corresponds to x/c = 0.30, 0.50, 0.60 and 0.70. The C p values at these points have been taken out of the training and testing datasets and are used as inputs to the ML model. The pressure distributions for both the training and testing datasets were averaged at 10 equispaced locations. The pressure values at the 4 input locations were not used in calculating these average pressure values. Figure 8 (a) shows the training data before it is averaged to 10 equispaced locations, and Figure 8 (b) shows the results after the data has been averaged. This averaged pressure can be considered as the reconstructed pressure distribution.

Figure 8: LHS. The training data used for M ∞ = 0.10 – 0.30, and α ∞ = -5.5 – 5.5

o . RHS. Red dots show the location of the 4 pressure values taken

as inputs to the DT model, with the pressure put into 10 bins that are equispaced.