Bearing fault diagnosis is inherently complex due to the varied interaction of fault parameters, the range of fault features, and the rapid increase in potential failure types. Obtaining training data for complex faults is often challenging in industrial settings, being sometimes unattainable, in contrast to the accessibility of single fault samples. To address limited training data for composite faults, this study introduces a semantic consistency embedding (SCE) zero-shot model for diagnosing composite faults. The model is trained using single fault samples to detect unseen composite faults. Validation with the CWRU dataset shows that the model achieves an 80.43% accuracy in classifying composite faults without any composite fault samples.