Predicting the consequences of accidents involving dangerous substances using machine learning

Application of rough set theory and machine learning algorithms in predicting accident outcomes in the Indian petroleum industry

“…These results indicate the efficiency of supervised ML approaches in predicting the accident outcome. The accuracy of our models shows superior performance to the existing studies in predicting accident consequences in the petroleum industry (Balasubramanian and Thangamani 41 ) accuracy of 0.92 using random forest and stochastic gradient descent (SGD), Kurian et al 42 achieved an accuracy of 0.948 using linear SVC RF for environmental consequences class, Chebila 45 highest accuracy achieved by RF 0.792 for human consequences class. Our Xgboost model has achieved 100% precision for near‐miss and catastrophic accident types of accidents, and MLP has achieved 100% precision for catastrophic accidents.…”

“…After performing an initial analysis using these ML techniques the overall metrics were not satisfactory, and then we used the smart grid parameter tuning method for finding the best parameters to get optimum performance of the models. 41,42,45 After measuring the predictive power of ML algorithms, we have used RST to explore the patterns behind the accident occurrence. The RST helps to understand the indiscernibility relationships between the given attributes under uncertain conditions even with noisy data.…”

“…Xgboost has achieved 94% precision for minor and major accident classes. The training accuracy of Xgboost and MLP is 0.95 and 0.94 overcoming the performance of the previous studies 41,42,45 . After measuring the predictive power of ML algorithms, we have used RST to explore the patterns behind the accident occurrence.…”

“…5,6 Class imbalance is also a significant issue that influences ML models' performance. Several authors 45,48 in the injury classification problem used synthetic minority oversampling technique (SMOTE), Borderline-MOTE, Majority Weighted-MOTE and so forth to under-sampling the majority class or oversample the minority class. These algorithms will generate a synthetic sample of minority classes using certain rules and patterns based on the input data structure.…”

“…This section presents a detailed analysis of studies that used ML techniques for analyzing and predicting workplace accidents in the global oil and gas industry. Chebila 45 used the eMARS database, which contains major accident reports from all European countries. A descriptive analysis of the accident database revealed that the majority (45%) of chemical industry accidents occur in petrochemical companies, primarily as a result of refining activities.…”