Structural health monitoring of mechanical assets can be hindered by environmental variability that causes distribution shifts between training and deployment. Many domain adaptation (DA) methods mitigate these shifts but behave as black boxes with limited insight into how representations change. This work introduces a novel interpretable framework, scattering-based prototype-aligned DA, that combines physics-guided feature extraction, synthetic data generation and prototype-based alignment for robust damage detection under temperature variation. A convolutional conditional variational autoencoder, trained on healthy data across temperatures with multi-domain reconstruction losses, generates temperature-conditioned synthetic damaged guided-wave signals from limited baseline damage measurements and healthy responses, creating a controlled testbed when damaged data at other temperatures are unavailable. Prototype-based domain adversarial training with gradient reversal and entropy-gated pseudo-labelling aligns source and target feature manifolds while preserving damage-sensitive patterns. Interpretability modules based on prototype trajectories, instance to prototype similarities and low-dimensional visualisations reveal how decision boundaries and latent representations evolve. Experiments on composite structures across temperatures show that the framework improves robustness over baselines and maintains high diagnostic accuracy while providing actionable insight into the adaptation process and enabling informed diagnostic assessment by domain experts in safety-critical contexts.
Rezazadeh et al. (Fri,) studied this question.