A Methodology for Feature Selection and Electrical Capability Prediction of a Coupled Shaker-DUT Model 75 (a) (b) (c) Fig. 6 VSD outputs of the three different spring stiffness coefficients plotted against the experimental ground truth. (a) Stiffness of 100N/m(b) Stiffness of 1000N/m(c) Stiffness of 10000N/m. The model failed to pick up the resonance that is occurring at ∼140 Hz and some of the oscillatory behaviors of the voltage quantity in the experimental ground truth. This is most likely a byproduct of the simplifications that were made in the lumped parameter model of the shaker, the simplification of the joint stiffness as being identical in all 3 directions, and also uncertainty in the Abaqus model of the BARC. The uncertainty in the FEM model of the BARC comes from the reduced order methods, indicated by the MAC diagonal values not being correlated to a value of one. However, qualitatively, this verifies a few components of the model. Firstly, the two systems, the Abaqus mass and stiffness and the lumped parameter model, are being coupled correctly. Secondly, the lumped parameter model that was calibrated using experimental data is sufficient to pick up the dynamic responses of the experimental system. Lastly, at higher stiffness values of the joint connections the Abaqus model of the BARC is tracking the dynamic responses of the experimental setup adequately. Interestingly, when visually inspecting the system at the fundamental frequency in Abaqus with a 100N/mstiffness value given for the springs, the top removable component separated from the box assembly base. This is most likely why there is less of a correlation to the ground truth in Figure 6a in comparison to the higher stiffness classes in Figures 6b and 6c. These distributions were then imported into the Bayes’ classifier and the minima were used to predict which of the three models best matched the ground truth. The classifier results shown in Table 3 predicted that the 10,000 N/mstiffness class most closely matched the experimental ground truth. The normalized probability of 1 shows that the joint connection of the physical system is best represented by a high stiffness value. Table3 Normalized and unnormalized probabilities output from the Bayesian classifier performed on three spring stiffness values. Class Unnormalized Bayes’ Prediction Normalized Bayes’ Prediction K value = 100 e−488385.36 0 K value = 1,000 e−3117.868 0 K value = 10,000 e−488.1526 1
RkJQdWJsaXNoZXIy MTMzNzEzMQ==