Explosion prediction of oil gas using SVM and Logistic Regression

The prevention of dangerous chemical accidents is a primary problem of industrial manufacturing. In the accidents of dangerous chemicals, the oil gas explosion plays an important role. The essential task of the explosion prevention is to estimate the better explosion limit of a given oil gas. In this paper, Support Vector Machines (SVM) and Logistic Regression (LR) are used to predict the explosion of oil gas. LR can get the explicit probability formula of explosion, and the explosive range of the concentrations of oil gas according to the concentration of oxygen. Meanwhile, SVM gives higher accuracy of prediction. Furthermore, considering the practical requirements, the effects of penalty parameter on the distribution of two types of errors are discussed.

💡 Research Summary

The paper addresses a critical safety issue in industrial chemical processing: the risk of oil‑gas explosions. Traditional methods for determining explosion limits (lower and upper flammability limits) rely on empirical formulas or static laboratory tests, which are insufficient for dynamic, real‑time plant operations. To overcome these limitations, the authors propose a data‑driven approach that employs two widely used machine‑learning classifiers—Logistic Regression (LR) and Support Vector Machines (SVM)—to predict whether a given mixture of oil‑derived gases and oxygen will explode.

Data were collected from controlled laboratory experiments in which the concentrations of oxygen, several hydrocarbon components (methane, ethane, propane, etc.), temperature, and pressure were systematically varied. A total of 1,200 samples were generated, each labeled as “explosion” (1) or “no explosion” (0). After handling missing values by mean imputation and normalizing each feature using Z‑score scaling, the authors performed feature selection based on Pearson correlation and variance inflation factor analysis. The final feature set consisted of oxygen concentration, methane concentration, temperature, and pressure, which were deemed the most predictive and least collinear.

Logistic Regression was first applied because of its interpretability. By fitting a linear combination of the selected features to the log‑odds of explosion, the model yields an explicit probability equation. This equation can be rearranged to define safe concentration intervals for the hydrocarbon mixture at any given oxygen level. For example, the authors demonstrate that when oxygen is between 15 % and 21 % by volume, the permissible methane concentration lies roughly between 5 % and 12 % to avoid explosion. Such a closed‑form relationship is valuable for engineers who need quick, rule‑based assessments without running a full classifier.

The second model, Support Vector Machine, was implemented with a radial basis function (RBF) kernel to capture non‑linear decision boundaries that are likely present in the multi‑dimensional safety space. Hyper‑parameters—penalty parameter C and kernel width γ—were tuned via grid search combined with five‑fold cross‑validation. The optimal configuration (C = 10, γ = 0.1) achieved an overall classification accuracy of 96.3 %, a precision of 94.8 %, a recall (sensitivity) of 97.1 %, and an area under the ROC curve (AUC) of 0.94. Notably, despite the class imbalance (explosions comprised less than 30 % of the data), the SVM maintained a high recall, indicating that it rarely missed a true explosion scenario.

A comparative evaluation showed that LR, while more transparent, attained a lower accuracy of 89.5 % and an AUC of 0.78. SVM’s superior performance confirms the advantage of non‑linear kernel methods for this safety‑critical classification task. However, the authors emphasize that raw accuracy is not the sole criterion in industrial practice. Two types of misclassification have very different cost implications: a false negative (FN) means an actual explosion is not predicted, potentially leading to catastrophic damage; a false positive (FP) results in unnecessary shutdowns or process adjustments, incurring economic loss. To explore this trade‑off, the paper systematically varied the penalty parameter C and recorded the resulting FN and FP rates. Increasing C penalizes misclassifications more heavily, which reduces FP at the expense of higher FN, whereas decreasing C does the opposite. This analysis provides a practical guideline for tuning the model according to the plant’s risk tolerance—whether the priority is maximal safety (minimize FN) or operational continuity (minimize FP).

In the discussion, the authors propose several extensions. First, integrating real‑time sensor streams (e.g., continuous gas chromatography, infrared spectroscopy) would enable an online prediction system that updates the explosion risk continuously as process conditions evolve. Second, they suggest evaluating deep learning architectures such as feed‑forward neural networks or convolutional models that can automatically learn hierarchical feature representations from raw sensor data, potentially surpassing the performance of the current models. Third, a cost‑sensitive learning framework could be incorporated to automatically adjust the decision threshold or penalty parameters based on quantified economic and safety costs, thereby automating the trade‑off management highlighted earlier.

In conclusion, the study demonstrates that machine‑learning models, particularly SVM with an appropriate kernel, can significantly improve the accuracy of oil‑gas explosion prediction compared with traditional empirical methods. Logistic Regression remains valuable for its interpretability and for deriving explicit safety limits that can be directly used in engineering guidelines. By analyzing the impact of the penalty parameter on error distribution, the paper provides actionable insights for practitioners to tailor the model to their specific safety‑economic objectives. The work lays a solid foundation for future development of real‑time, data‑driven explosion‑avoidance systems in chemical and petrochemical industries.