Partial wear of Tunnel Boring Machine (TBM) disc cutters reduces rock-breaking efficiency and increases the risk of unplanned downtime; therefore, accurate state identification is essential for safe and efficient TBM tunnelling. However, disc-cutter vibration signals are strongly affected by changes in rotational speed, load, and cutter–rock contact conditions, resulting in substantial distribution shifts across operating conditions. Moreover, partial-wear samples generally rely on shutdown inspection for labelling, leading to insufficient labelled data under target operating conditions and limiting the cross-condition application of deep learning models. To address these issues, an unsupervised domain adaptation method, termed WSMCD, is proposed for partial-wear state identification of TBM disc cutters. Using a self-developed disc-cutter rock-breaking test rig, multi-cutter coupled rock-breaking conditions with different rotational speeds and loads were established, and multichannel vibration signals of normal and eccentrically worn disc cutters were collected to construct a cross-condition transfer dataset. Experimental results show that WSMCD achieves an average accuracy of 98.64% across six cross-condition transfer tasks. Further quantitative analysis using the proposed normalized cross-domain shift index and target-domain inter-class-to-intra-class discriminability ratio demonstrates that the Wasserstein distance plays a primary role in constraining cross-domain feature shifts and alleviating negative transfer, while the SE module further improves the class-discriminative structure of target-domain features under the Wasserstein constraint.