Effective fault diagnosis is a key link in ensuring the safe and stable operation of industrial processes and equipment. Currently, China is undergoing a pivotal transition from an industrial powerhouse to an industrial leader. During this process, smart manufacturing, characterized by digitalization and intelligence, is accelerating, serving as a key support for industrial transformation and upgrading. As intelligent manufacturing advances, industrial processes are becoming increasingly complex, and fault types are becoming more diverse, which imposes higher demands on diagnosis methods. Conventional supervised learning-based fault diagnosis methods depend heavily on substantial labeled samples to train models, focusing on classifying faults that have encountered during training (seen faults). In real scenarios, industrial fault diagnosis often faces challenges such as incomplete fault categories and significant distribution discrepancies. Zero-sample fault diagnosis, which can diagnose unseen fault types without target fault training samples, has emerged as a crucial research direction to address these issues. First, zero-shot scenarios are categorized into traditional and generalized zero-shot settings. After reviewing the development of industrial fault diagnosis techniques, the concepts and connotation of traditional and generalized zero-sample fault diagnosis are combed, and their differences are clarified from conventional methods. Second, the current research on zero-shot fault diagnosis is systematically surveyed from perspectives such as auxiliary knowledge construction, methodological implementation, and application scenarios. Furthermore, key challenges and potential research extensions are discussed for existing approaches. Finally, representative datasets and open-source code are compiled, while future trends and challenges are outlined to provide theoretical references and technical support for zero-sample fault diagnosis.