Zero‐Shot‐Motivated Intelligent Fault Diagnosis: A Survey of Methods, Applications, and Future Directions

Conventional fault diagnosis (FD) methods depend heavily on model precision and data quality, while traditional machine learning and deep learning pipelines typically require large volumes of labelled data. The development of transfer learning has opened an alternative route for learning‐based FD. In particular, zero‐shot‐motivated fault diagnosis seeks to recognise unseen faults under scarce or missing labels by exploiting priors rather than exhaustive annotation, and has attracted increasing attention in industrial monitoring. Despite rapid progress, the field still lacks scenario‐aligned evaluation and reporting standards, and a comprehensive survey is still missing. This article presents a structured review of zero‐shot‐motivated FD methods to summarise the state of the art and support the design of reliable solutions. It first introduces the theoretical background of zero‐shot learning and organises semantic and non‐semantic priors together with their operational mechanisms. It then surveys representative methods by application scenario and compares evaluation goals and metrics across scenarios. Practical guidelines for method selection towards deployment are provided, followed by a discussion of key challenges and future research directions. The survey concludes with a synthesis of key takeaways and aims to promote standardised evaluation and engineering adoption of zero‐shot‐motivated FD.