Prediction of explosion hazard of aluminum powder two-phase mixed system using random forest based on K-fold cross-validation

Abstract

Dust explosion is a significant safety hazard in various industries and trades, with aluminum dust being particularly sensitive to ignition and leading to severe explosion consequences. During the actual industrial production process, there is a risk of aluminum powder dust coming into contact with other metal dust, flammable and explosive gases and liquids. Aluminum dust and other flammable and explosive substances, even if the concentration of aluminum dust and other flammable and explosive substances are below the lower explosive limit, the aluminum dust as the dominant multi-phase mixing system still has the potential to explode. Compared to a single aluminum dust, the aluminum powder multiphase system has a higher explosion hazard. Thus, the rapid prediction of the explosion intensity of the gas-liquid-solid multiphase system with aluminum dust as the main body is of great significance for the assessment of the explosion hazard of the mixed dust. On the other hand, the explosion properties of the gas-liquid-solid multiphase system with aluminum powder as the main body are affected by a number of factors such as powder particle size, powder concentration, combustible material concentration, and different systems have a large variability between. This is a significant challenging for the rapid and precise prediction of the explosive intensity of aluminum powders in multiphase systems. In this study, machine learning methods (random forest (RF) and multilayer perceptron (MLP)) were applied to deeply excavate the nonlinear relationship between the explosion index (K_st) of gas-liquid-solid multiphase system with aluminum dust as the main body and the explosion influencing factors. Feature engineering was employed during the model building process for improving the data representation model. The grid search method, including the K-fold cross-validation and three-model performance evaluation metrics, were incorporated to optimize, assess, and test the model's state and performance. A total of 233 K_st samples were gathered, with 163 samples (70% of the total samples) were allocated for training while the other 70 samples (30% of the total samples) were adopted for testing. By adopting the same dataset, compared with the MLP model, the RF model exhibits enhanced generalization capability and higher prediction accuracy, with a prediction accuracy of about 90%. Furthermore, the gas-solid two-phase dominated by aluminum powder had the highest prediction accuracy in the RF model, followed by liquid-solid and solid-solid systems. This study could help to rapidly predict the explosion intensity of the multi-phase system of aluminum powder occurring under complex conditions, and paved a way for decision-making of multi-factor affected emergency.