PSI - Issue 80

Nima Rezazadeh et al. / Procedia Structural Integrity 80 (2026) 411–417 Author name / Structural Integrity Procedia 00 (2019) 000 – 000

415

5

0.5 and 0.1. Training ran for 50 epochs with Adam (learning rate = 0.001). Target- domain data was split 85 % for adaptation, 15 % for testing. Fig . 3-a and 3-b show confusion matrices of the target-domain fault diagnosis.

(a)

(b)

Fig. 3. Confusion matrices of the test phase of the target domains for (a) Paderborn dataset and (b) CWRU dataset

The results shown in Fig. 3-a and 3-b demonstrate the robustness of Proto-ADDA-X in domain adaptation and subsequent fault diagnosis, with both datasets achieving 100% classification accuracy even though the target domain labels remained unknown throughout the adaptation process. Fig. 4-a and 4-b show, respectively, the feature and prototype distribution and the prototype trajectories across training for the Paderborn dataset; Fig. 4-c and 4-d present the corresponding plots for the CWRU dataset.

(a)

(b)