Unexpected failures in rotating machinery cause substantial economic losses and safety hazards in industrial systems, underscoring the practical need for reliable condition-based monitoring. Intelligent fault diagnosis must simultaneously address three real-world constraints in Industrial IoT deployments: severe data scarcity, dynamic distribution shifts, and the presence of previously unseen fault categories. This paper reviews a decade of methodological progress (2016–2026), tracing the development from unsupervised clustering and few-shot meta-learning to open-set single-source domain generalization. We propose a unified six-tier taxonomy classifying methods into data-level synthesis, metric/meta-learning, domain adaptation, domain generalization, open-set recognition, and an emerging category covering attention-based and self-supervised approaches. We examine the integration of wavelet-guided generative models with adversarial and prototype-aware alignment objectives, contextualize convergent findings on attention mechanisms and self-supervised pre-training that corroborate the role of structural frequency priors, and show that purely distributional alignment is insufficient under open-set conditions without latent-space structural constraints. Five open problems are identified: OS-SSDG, calibrated conformal thresholding, physics-informed class completion, zero-shot disentanglement, and foundation-model pre-training for IIoT. We conclude that Conformal Prediction combined with wavelet-structured representations offers a principled path toward bounded, interpretable fault diagnosis under realistic industrial constraints.