PSI - Issue 80

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

414

4

2.6. Explainability outputs To enhance interpretability of the proposed domain-adaptation framework, two complementary visualization methods were used. The first shows feature and prototype distributions: a two-dimensional PCA scatter of all source and target embeddings at a selected epoch, with points colored by class and styled by domain, and class prototypes overlaid as distinct markers. This view illustrates target-source alignment and feature clustering by class. The second shows prototype trajectories: PCA-projected paths of source and target prototypes across epochs, with successive centroids linked by arrows to reveal class-level convergence and remaining domain misalignment during training. 3. Case studies In this work, to validate the designed framework, 2 experimental bearing datasets were considered that their brief details can be found in the following. 3.1. Paderborn university bearing fault benchmark The Paderborn University Bearing Fault Benchmark (Lessmeier et al., 2016) records vibration from a drive-train test rig under four load-speed settings (Fig. 1); here, Condition 1 (source) runs at 1500 rpm, 0.7 Nm and 1000 N, whereas Condition 2 (target) runs at 900 rpm with the same torque and radial force. Each condition yields 20 samples for four states: healthy, outer-race fault, inner-race fault and combined inner-outer fault, produced by real fatigue on identical 6205-steel bearings K001, KA04, KI04 and KB23, with material and geometry unchanged across domains. Only vibration signals are analysed, and labels 0, 1, 2 and 3 map to healthy, outer, inner and combined.

Fig. 1. Paderborn dataset test rig

3.2. Case Western Reserve University (CWRU) Bearing Fault dataset The CWRU dataset (Loparo, 2012) contains fan- and drive-end vibration from a motor rig (Fig. 2); we use the 48 kHz drive-end signals, also at 12 kHz. Each recording is split into twenty windows. The source domain is 1 HP at 1772 rpm, the target 3 HP at 1750 rpm. Four states: healthy, inner-race fault, rolling-element defect, and an outer-race fault at 3 o’clock, 90° from load zone share identical bearings and acquisition settings. Labels are 0 normal, 1 inner, 2 ball, 3 outer.

Fig. 2. CWRU dataset test rig

4. Results The Proto-ADDA-X framework was tested on the Paderborn and CWRU datasets. Feature extraction used 13 MFCCs and the “db4” wavelet; sampling rates were 16 kHz and 48 kHz, respectively. A two -layer MLP with 128 ReLU units and 0.3 dropout handled classification. Prototype- attention used α = 0.1 and τ = 0.6. Adversarial losses and were