Application of an explainable glass-box machine learning approach for prognostic analysis of a biogas-powered small agriculture engine

Biogas has developed as a potential substitute fuel source due to its renewable and sustainable nature, which can help reduce greenhouse gas emissions. In this paper, along with the experimental investigation, a novel interpretable expert system was constructed to address these issues and provide meaningful explanations for model predictions using extreme gradient boosting (XGB). For this purpose, the SHapley Additive exPlanations (SHAP) tool was coupled with XGB open-source algorithm to simulate five efficiency parameters of the biogas-powered engines, including brake thermal efficiency (BTE), peak pressure (PP), hydrocarbon (HC), oxides of nitrogen (NOx), and carbon monoxide (CO). Here, four experimental-based variables comprised of fuel injection timing, fuel injection pressure, compression ratio, and engine load were employed as predictors. Apart from the main framework, two advanced ensemble machine learning (ML), namely the light gradient-boosting machine (LightGBM) and Extra Tree algorithms, were adopted to validate the primary model. In addition, we use the SHAP framework to understand the impact of input features on engine performance and emission outputs. XGB-SHAP owing to its best predictive performance (BTE|R = 0.994 and RMSE = 0.567, PP|R = 0.984 and RMSE = 0.846, HC|R = 0.994 and RMSE = 6.215, NOx|R = 0.998 and RMSE = 1.407, and CO|R = 0.985 and RMSE = 3.464) outperformed the Extra Tree and LightGBM, respectively.