Predictive Optimization of Vibrations in a Biodiesel-Fueled Variable Compression Ratio Engine Using BO-CNN-LSTM

Abstract Integrating sustainable biodiesel blends into variable compression ratio (VCR) diesel engines significantly alters combustion dynamics and structural vibrations. The present study introduces a Bayesian optimized 1D convolutional neural network-long-short-term memory (CNN-LSTM) deep learning framework designed for optimizing the multiaxis engine vibrations. The predictive architecture processes a six-parameter input matrix encompassing engine load, compression ratio (CR), and critical thermophysical fuel properties. A precise evaluation demonstrated the superior fidelity of the hybrid network, achieving up to a coefficient of determination (R2) of 0.985 and a root mean square error (RMSE) of 0.288 m/s2, substantially outperforming optimized support vector machine (SVM) and multilayer perceptron (MLP) benchmarks. Multiaxis spatial stability was visually confirmed using Taylor diagrams, while a 3000-iteration bootstrap analysis validated algorithmic robustness within a narrow 95% confidence interval (0.38–0.79 m/s2). Permutation feature importance and SHapley Additive exPlanations (SHAP) analysis clarified the thermodynamic decision-making process, revealing that elevated engine load and CR intensify mechanical ringing, whereas the higher density and viscosity of biodiesel blends act as natural vibration dampeners. High-resolution contour maps generated through systematic response surface mapping provided precise predictive optimization surfaces. Deploying this advanced methodology establishes a highly reliable, cost-effective virtual calibration tool for next-generation biodiesel-fueled engines, eliminating the necessity for exhaustive physical testing.