Abstract: In water conservancy and hydropower projects, construction safety hazards come large in volume and are usually diversified in forms and types. The same hidden danger may involve multiple types, but the definition of its types is ambiguous; previous classification of hazard types is mostly dominated by manual experience, easily leading to confusion and difficulty in hazard management. To address such issues, this paper presents a multi-label intelligent identification method for construction safety hazards in hydropower projects. First, the ALBERT model is used to encode text information to achieve a high-precision quantification of the unstructured risk texts. Then, we construct a multi-label intelligent classification model of safety hazards text to improve identification efficiency, considering the content weight of the Chinese text for safety hazards and using the Bidirectional Gated Recurrent Unit (Bi-GRU) improved by the Attention mechanism. Finally, the performance of this method is tested using the texts of hydropower engineering construction safety hazards, verifying its F1 value reaches 92.11% compared with previous text classification methods. It is proved applicable, as a useful support to safety management of hydropower construction and analysis of its safety hazards.

Key words: hydropower engineering construction, safety hazard, text multi-label intelligent classification, ALBERT, bidirectional-GRU, attention mechanism