Hierarchy-aware prompt tuning with bidirectional information propagation for failure data mining

Failure-related textual data contain rich hierarchical information essential for understanding system states and supporting Operation and Maintenance (O&M) decision making. Hierarchical text classification is central to mining such data and can reduce costs and improve efficiency in the O&M activities of complex engineering systems. Yet the intricacy of failure narratives often yields suboptimal performance, especially in low-resource, few-shot settings. This paper presents a hierarchy-aware prompt tuning approach built on Pre-trained Language Models (PLMs) for few-shot hierarchical classification of failure texts. At its core is a Bidirectional Layer-wise Propagation (BLP) mechanism that injects label structure information during model learning, together with an uncertainty-quantification scheme that enhances robustness and practical applicability in O&M workflows. The proposed approach is designed to integrate seamlessly with different types of PLMs, including BERT, GPT, T5, etc. Validated on a large-scale offshore wind-turbine O&M corpus from 313 units and additionally on a public DBpedia corpus with a deeper and wider taxonomy, the approach delivers strong overall performance and remains effective under extremely low-resource conditions. These findings indicate that the framework offers a robust, data-efficient solution for failure text mining in offshore wind and a promising candidate for other industrial O&M settings.