A single-domain generalization fault diagnosis method for rolling bearings based on Kalman state embedding and uncertainty-aware tokenization

Single-domain generalization fault diagnosis aims to train a model using labelled data from a single source domain and recognize faults in unseen target domains without accessing target-domain data during training. This task remains challenging because vibration signals are non-stationary and vary significantly across operating conditions. To address this problem, this paper proposes SKUT-BERT, a fault diagnosis framework that combines dynamic-routing multi-expert Kalman state embedding with uncertainty-aware multi-scale tokenization. The proposed method directly models raw one-dimensional vibration signals without image or text conversion. The Kalman state embedding module captures local state evolution, innovations, and uncertainty information, while the uncertainty-aware tokenization module adaptively integrates multi-scale temporal features to generate discriminative signal tokens. These tokens are then fed into a BERT encoder for long-range dependency modelling and fault classification. Experiments are conducted on three bearing datasets, namely SDUST, JNU, and CWRU, under six single-source-to-unseen-target diagnostic tasks for each dataset. SKUT-BERT achieves average accuracies of 94.41%, 96.37%, and 99.93% on the three datasets, respectively, improving over the strongest baseline by 2.33, 4.92, and 0.93 percentage points. These results demonstrate its effectiveness for single-domain generalization fault diagnosis.