40 S. Carter validation DOF set that could be used for a CV check. One training set contained “good” response DOFs that describe the response deflection shapes well, and the other training set contained “bad” response DOFs that poorly describe the response deflection shapes. Consequently, it is assumed that the validation DOF set will not add independent information to the ISE problem when the good DOFs are used in the training set (since the deflection shapes are well defined by the good DOFs) but will add independent information when the bad DOFs are used in the training set (since the bad DOFs poorly describe the deflection shapes). As a result, the different DOF sets make it possible to test the previously described hypothesis on CV checks, depending on if the validation DOFs can identify overfitting when the sources are estimated with responses at the good and bad DOFs (where comparisons between the estimated and truth sources are the ultimate check). The DOF sets were objectively selected with an effective independence (EFI) [28] algorithm to eliminate engineering bias from the results of the study. Both the good and bad training DOF sets used thirty DOFs at ten triaxial response locations and the validation DOF set used twelve DOFs at four triaxial response locations. Note that the same validation DOF set was used regardless of the training DOF set. The good DOF set was selected with the typical EFI process and metric for triaxial accelerometers, the selected (good) locations are shown in Figure 4. The bad DOF set also used the typical EFI process for triaxial accelerometers, but the EFI metric was inverted so the worst sensor locations were chosen, the selected (bad) locations are shown in Figure 5. The validation DOF set was also selected with the standard EFI process and metric (for triaxial accelerometers), but the algorithm was initialized with the good and bad DOF sets, the selected (validation) locations are shown in Figure 6. The “truth” target responses for the overfitting study were computed by applying broadband random excitation to triaxial input DOFs at the corners of the frame (twelve DOFs total) in the truth model. The input DOFs are shown in Figure 7. Fig. 4 The good set of response DOFs for the overfitting study. Fig. 5 The bad set of response DOFs for the overfitting study.
RkJQdWJsaXNoZXIy MTMzNzEzMQ==