Ensemble machine learning models of a compliant seat coupled with the occupant for predicting the seat transmissibility with dual-axis vibration

Improving riding comfort requires models to precisely predict the dynamic characteristics of a compliant seat coupled with the occupant, but it has been challenging due to the complex dependence on anthropometric parameters and vibration conditions. Most of the machine learning (ML) models developed so far have been confined to predicting biodynamic responses, and further enhancements in both accuracy and interpretability are still required. In the present study, ML models were utilized to predict the vertical in-line and fore-and-aft cross-axis transmissibilities of the cushion during different levels of dual-axis vibrations between 1 and 10 Hz. Specifically, the dataset was established through whole-body vibration experiments, where 12 participants were exposed to vertical excitation at three different amplitudes with and without additional fore-and-aft excitation. Two standalone ML models (Artificial Neural Network (ANN), Support Vector Machine Regression (SVR)), and three ensemble ML models (Random Forest (RF), eXtreme Gradient Boosting (XGBoost) and Light Gradient Boosting (LightGBM)) were developed, with input variables including the frequency and amplitude of vibration excitation, the age, height, weight, buttock width, and knee height of the subjects. The ensemble model XGBoost outperformed the others, demonstrating the highest accuracy. Moreover, the SHapely Additive exPlanations (SHAP) method was used to interpret the prediction of XGBoost model, revealing that the frequency and amplitude of the excitation were the key variables affecting the transmissibilities. This study suggests that the combination of SHAP and XGBoost shows promising potential for predicting seat transmissibility under the tested vertical and fore-aft vibration excitation conditions and improving the interpretability of ML models for seating and suspension design.