Research on risk assessment of hot work based on deep learning

Deficiencies in existing research on hot work risk assessment were identified, as work number and time factors were not considered. A risk assessment method integrating these two factors was proposed. First, a time-segmented adjacent mean method was designed to handle outliers. A time-segmented Lagrange interpolation method was constructed to process missing values, so as to improve data quality. Second, a hot work quantity prediction model based on Attention-enhanced Bidirectional Long Short-Term Memory (Attention-BiLSTM) neural network was built. It achieved accurate prediction of hot work quantity. Experiments show that the accuracy of this model reaches over 90% and is superior to models such as RNN and GRU. Finally, a risk assessment model coupling quantity and time factors was established. By introducing work level weights and time weights, hot work risk values at different times were calculated. Experimental results confirm that this method can effectively reflect risk differences in different time periods, realize rapid and accurate assessment of hot work risks, and provide strong support for the safety management and control of hot work in chemical enterprises.