Research on upstream design support using machine learning (Extraction of pleasant factors from engine and electric motorcycle driving data)

The purpose of this study is to investigate a machine learning method to support upstream of design process by taking a motorcycle design targeting riding comfort as an example. In our previous report, a machine learning decision tree model was trained using temporal factors with the riders’ subjective evaluation as the target variable while the physical and dynamic states of the bike and the riders were used as predictors. In the present study, a high performance deep-learning regression neural network was trained to predict the real-time subjective evaluation of riders using physical state data captured over a certain time period. Deep learning algorithms, however, often yield black box models which are inadequate for the purpose of acquiring design knowledge. Hence, the feature importances of the models were extracted using explainable artificial intelligence (XAI). Relationships between physical states and changes in rider’s pleasure were elucidated by applying XAI on the deep learning model and bike riding data. As the result, two design guidelines to increase riding comfort were obtained as (1) improve the throttle feel and responsiveness, and (2) improve vibration control at the rear of the motorcycle and rear suspension.