Research on emulsion concentration detection technology based on interpretable machine learning methods

• The study is based on 189 sets of field dispensing experiments, and the experimental data are used as a dataset for machine learning training, which is validated with five machine learning methods after dividing the dataset and test set, demonstrate RF model's superior performance with lowest errors (MSE=0.115, MAE=0.08) and highest R² (0.98) among five ML algorithms through grid search optimization. • Identify EC² as dominant predictor via SHAP analysis, revealing positive correlation between emulsion concentration and conductivity ( p < 0.001). • Validate practical effectiveness through IoT deployment showing 4.02 % mean relative error, establishing reliable industrial monitoring system. Emulsion plays a critical role in ensuring the normal operation of hydraulic equipment and coal mine safety. However, existing underground emulsion concentration detection modules suffer from high failure rates and insufficient accuracy. With the increasing demand for intelligent production systems, there is an urgent need for a precise, easily deployable, and highly intelligent emulsion concentration detection solution. To address these challenges, this study developed a cloud-based IoT system utilizing machine learning (ML) for emulsion concentration detection. Laboratory experiments were conducted to establish a dataset correlating emulsion concentration with liquid temperature and electrical conductivity. Five ML models—XGBoost, Random Forest (RF), Linear Regression, Support Vector Regression (SVR), and Ridge Regression—were trained and hyperparameter-optimized. Validation on training and test datasets demonstrated that the RF model outperformed others, achieving the lowest MSE (0.115), MAE (0.08), RMSE (0.339), MRE (0.011), and MAPE (1.164), along with the highest R² (0.981). Taylor diagram analysis further confirmed the RF model’s superior alignment with observational data. SHAP-based interpretability analysis identified electrical conductivity as the primary predictive factor, with temperature exhibiting minimal influence. The optimized RF model was deployed on a cloud platform integrated with a Siemens S7–200SMART PLC via the Aprus-X IoT framework. Field validation through 20 emulsion preparation trials revealed an average relative error of 4.02 %, demonstrating the system’s reliability and precision for industrial applications.