Research on risk assessment of hot work based on deep learning

Area of Science:

• Industrial Safety
• Risk Management
• Artificial Intelligence in Engineering

Background:

• Current hot work risk assessment methods are insufficient due to the omission of critical factors like work volume and temporal variations.
• This gap necessitates the development of advanced methodologies for more accurate and dynamic risk evaluation.

Purpose of the Study:

• To propose a novel hot work risk assessment method that integrates work quantity and time factors.
• To enhance the accuracy and timeliness of risk assessment for improved safety management in chemical enterprises.

Main Methods:

• Data preprocessing techniques including time-segmented adjacent mean for outlier handling and time-segmented Lagrange interpolation for missing value imputation.
• Development of a hot work quantity prediction model utilizing an Attention-enhanced Bidirectional Long Short-Term Memory (Attention-BiLSTM) neural network.
• Establishment of a risk assessment model that couples predicted hot work quantity with time-dependent risk factors, incorporating work level and time weights.

Main Results:

• The Attention-BiLSTM model achieved over 90% accuracy in predicting hot work quantity, outperforming traditional models like RNN and GRU.
• The proposed risk assessment method effectively differentiates risk levels across various time periods.
• Experimental validation confirmed the model's capability for rapid and precise hot work risk assessment.

Conclusions:

• The integrated approach significantly improves hot work risk assessment accuracy by incorporating work quantity and time dynamics.
• This method provides a robust framework for proactive safety management and control of hot work activities in industrial settings, particularly chemical enterprises.
• The study highlights the potential of advanced machine learning techniques in enhancing occupational safety and risk mitigation strategies.