Machine Learning-Based Monitoring of Chemical Contamination from Drilling Leaks

With the increasing requirements for environmental protection in oil and gas field development, the leakage during drilling and the chemical pollution caused by it have become the key technical difficulties restricting sustainable development. In this study, a set of intelligent monitoring and risk assessment system integrating multi-source data is constructed, based on three uncertainty quantification methods, namely, Dropout Bayesian Approximation, Density Network Integration, and Mixed Density Network (MDN), to deeply mine and distribute the drilling parameter and chemical detection data of 17 wells. To address the limitations of the traditional methods, which are difficult to take into account the nonlinearity, multiple peaks and uncertainty, SHAP value and BPCA are introduced for feature selection and dimensionality reduction, and the MDN model is optimised based on maximum likelihood and entropy regularisation strategies. The experimental results show that the MDN outperforms other methods in terms of negative log-likelihood (NLL), mean square error (MSE) and confidence interval coverage (PICP), and achieves accurate fitting of the pollution concentration distribution and risk warning. Finally, combining the pollution probability distribution and well control scheduling rules output from the model, the parameter optimisation and operation recommendation scheme is proposed, which achieves an early warning accuracy of more than 93% in the field deployment, and provides strong support for green drilling and environmental protection.