Assessing the health of fast-switching electrical contacts is challenging due to high-dimensional cycle-resolved measurements and the limited expressiveness of hand-crafted diagnostic indicators.
In this thesis, a variational autoencoder is trained on multi-channel recordings of voltage, current, and contact position, covering tens of thousands of make-break cycles, to compress contact behavior into a compact latent representation.
To ensure interpretability, the learned latent dimensions are explicitly aligned with established domain features and coupled through a bidirectional generative mapping between physical descriptors and latent codes. The proposed approach is benchmarked against foundation-model and Bayesian neural network baselines to assess robustness and uncertainty awareness.
The results establish an explainable link between data-driven embeddings and physically meaningful features, enabling a critical re-evaluation of legacy indicators for contact-state diagnosis.