Intelligent identification of hydropower engineering safety hazards: research on text classification method based on deep learning

To achieve timely identification and prevention of potential construction risks, this paper presents an intelligent classification approach based on deep learning. Given that such texts abound with technical terms and have a complex engineering background, traditional classification techniques struggle to effectively capture their profound features, thereby leading to limited classification accuracy. First, the ALBERT model is employed to extract the dynamic semantic representation of the text, and subsequently, the complex semantic associations between the text are deeply captured through the multi-layer Long Short-Term Memory network (LSTM) coding layer. In the decoding stage, multi-layer LSTM integrated with the attention mechanism is introduced to enhance the inter-label dependency and optimize the multi-label sequence prediction. The experiment was carried out based on a disaster text dataset of a hydropower station, and the results indicated that the F1 score was as high as 90.64%, significantly enhancing the classification efficiency and safety management efficiency, and offering robust support for the formulation of precise preventive measures.