Abstract: Long-distance water transmission project lines, complex environment along the line, in the daily operation of water transmission projects, engineering safety inspection is an important way to maintain production safety. A large amount of inspection text data is generated in the engineering inspection. In the traditional production management process, inspection text relies on managers to manually classify according to the severity of the problem, which is inefficient and prone to subjective problem classification errors, and is not sufficient for good management of long lines and complex environment along the water transmission project. To address this problem, this paper proposes a hybrid deep learning model combining Bi-directional Long Short-Term Memory（BERT） and BERT neural network to classify inspection text intelligently by using BERT as the input layer to transform inspection text into feature vectors, and then inputting the results to BiLSTM.The model uses BERT as the input layer to transform inspection text into feature vectors, and then feeds the results into BiLSTM model to mine text features and realize intelligent classification of inspection text. Compared with the mainstream deep learning models TextCNN, BERT and BiLSTM, the accuracy, recall and F1 values of the model reach 92.30%, 92.32% and 92.30% respectively, and the model outperforms other deep learning models.