Abstract: Knowledge mining based on hidden danger troubleshooting information plays an important role in supporting engineering safety management; the natural language processing (NLP) technology is an important method to realize text knowledge mining. The depth and accuracy of knowledge mining are the key indicators of such methods. This paper presents a new hidden hazard text knowledge mining method that combines text classification and text mining technology to improve its efficiency in application to hydropower projects. It uses the RoBERTa-wwm-CNN hybrid deep learning model to make a fast intelligent classification of hidden hazard texts. On this basis, it realizes a visual analysis of the key points of hidden danger management through drawing a nephogram for hidden hazard words, and analyzes inner links among latent danger data via constructing a word co-occurrence network. Application to a hydropower station for comparison with the existing advanced text classification models shows that our new model is better in accuracy and applicability.

Key words: hydropower engineering, safety hazard, knowledge mining, text classification