In the task of fault diagnosis, information extraction (IE) serves as a key technology in the data preparation process. However, the operating environment of aviation equipment (e.g., aircraft) is complex and highly variable, and new types of fault patterns keep emerging, making it extremely difficult to obtain comprehensive and accurate labeled data for traditional supervised methods. Large language models (LLMs) pre-trained on massive corpora have acquired extensive prior knowledge and powerful zero-shot learning abilities, presenting significant opportunities for IE tasks in low-resource scenarios. Zero-shot IE, especially zero-shot named entity recognition (NER), still faces some challenges, including the failure to fully capture the unique features of NER tasks, oversight of correlations between contexts surrounding entities, and indiscriminate utilization of irrelevant task demonstrations. This paper presents a self-enhancement framework to better address these issues. Specifically, we propose a two-stage self-annotation strategy for unlabeled corpus to obtain pseudo in-context learning (ICL) demonstrations. Additionally, use mutual information criteria to extract class-level type-related features (TRFs) and take them to retrieve relevant ICL demonstrations. The results of ablation experiments show that the proposed method is effective for zero-shot NER tasks based on large language models.