A Cross-Condition Fault Diagnosis Method for Rolling Bearings Combining Content-Conditioned Positional Encoding and Label-Conditioned In-Context Learning

Abstract Cross-condition fault diagnosis for rolling bearings is still a difficult task, mainly because vibration signals acquired under varying operating conditions exhibit distribution differences. To address this issue, this paper proposes a cross-condition fault diagnosis method named CCPE-LCIC, which combines Content-Conditioned Positional Encoding and Label-Conditioned In-Context Learning. The proposed method employs a Transformer-based encoder to extract fault-related features from raw vibration signals, while adaptively enhancing temporal position information according to signal content. Meanwhile, label information from support samples is incorporated into the in-context learning process to improve the modeling of feature-label relationships and support query classification. An auxiliary supervised loss is further introduced to enhance feature discriminability and stabilize training. The effectiveness of the proposed method is verified through cross-condition experiments conducted on the CWRU and TBVD bearing datasets. Ablation studies further confirm the contributions of the main components.