106 E. M. Tronci et al. -5 0 5 LDA Dimension 1 (a) -6 -4 -2 0 2 4 6 D U -6 -4 -2 0 2 4 6 LDA Dimension 3 (b) -6 -4 -2 0 2 4 6 LDA Dimension 4 D U -5 0 5 LDA Dimension 1 (a) -6 -4 -2 0 2 4 6 LDA Dimension 2 D U -6 -4 -2 0 2 4 6 LDA Dimension 3 (b) -6 -4 -2 0 2 4 6 LDA Dimension 4 D U Fig. 6 First four components of x-vectors transformed via LDA for the binary classification task considering the noise augmented VoxCeleb dataset in case B: (a) LDA Dimension 1 vs. LDA Dimension 2; (b) LDA Dimension 3 vs. LDA Dimension 4 -5 10 0 0 LDA Dimension 3 5 LDA Dimension 1 LDA Dimension 2 -10 0 10 -20 -10 D1 D2 D3 D4 D5 D6 U (a) -5 0 10 0 LDA Dimension 3 5 LDA Dimension 2 LDA Dimension 1 -10 0 10 -20 -10 D1 D2 D3 D4 D5 D6 U (b) Fig. 7 Three components of x-vectors transformed via LDA for the multiclass by damage type classification task considering the noise augmented VoxCeleb dataset in case B: (a) LDA Dimension 1 vs. LDA Dimension 2 for damaged and undamaged scenarios; (b) LDA Dimension 1 vs. LDA Dimension 2 for damaged scenarios the MUSAN dataset is considered for augmenting the dataset (Fig. 5). Figure 7 shows the LDA-transformed x-vectors for the three LDA dimensions in the case of noise augmented VoxCeleb dataset. Intuitively, the x-vectors features linked with the most severe damages, scenarios D1 and D2, are perfectly separated from the clusters describing the other damaged and undamaged scenarios. The clusters representing the features related to damage type D4 and D5, which are less severe failure mechanisms with respect to cases D1 and D2 and consist of more localized damage, are still correctly classified. Damage D3, which still interests the loss of stiffness of a whole structural element, is correctly distinguished from the undamaged conditions, even with some wrongly classified cases. On the other hand, damage type D6, which corresponds to a partial loss of stiffness for one structural element, is improperly classified.
RkJQdWJsaXNoZXIy MTMzNzEzMQ==