Abstract: To enhance the investigation and management of potential hazards in hydropower construction, workers can use mobile reporting to announce safety hazards promptly. However, hazard classification and its accuracy are often subjective, and manual correction is time-consuming and labor-intensive. To mitigate confusion in hazard management during construction, this paper describes a NRBO-CNN-BiLSTM method for auxiliary correction of the mobile phone-reported hazard categories. First, safety hazard data are tokenized, preprocessed, and converted into word vectors, followed by normalization. Then, we apply an attention mechanism to enhance the feature representation capability, and construct a safety hazard classification model using convolutional neural networks and bidirectional long-short-term memory networks. Finally, we work out a Newton-Raphson optimization algorithm to train the model for optimal parameters selection. Case studies demonstrate the probability is 69.2% for the classification of 18 types of hazards. The main reason lies in a relatively low frequency of certain hidden danger categories. In the tests of 6 hazard categories with balanced datasets, our new model achieves a classification probability of 94.6%, a recall value of 94.6%, and an F1 score of 94.6%. The accuracies of these indexes are superior to those of alternative classification models, indicating this correction model is effective and better.

Key words: hydropower project, snapshoot, construction safety, potential safety hazards, text classification