Transfer Learning Utilizing Real Industrial Vehicle Data Based on Similar Usage Trends

This study proposed a lifespan prediction method for fuel-cell systems in industrial vehicles within the framework of transfer learning. Since industrial vehicles with fuel cells have only recently become widely used, there is a dearth of data per vehicle. Additionally, the use of such vehicles differs among customers. Under these constraints, the model prediction by ridge regression showed that some vehicles missed the prediction significantly. Therefore, to achieve more consistent performance across vehicles, we considered a transfer learning method for applying knowledge from the data of other vehicles. Specifically, as an extension of ridge regression, we propose a regularization term that transfers knowledge from other vehicle data and grouping method for creating a group of similar vehicles. Validations on a real dataset of vehicles in operation showed that the proposed method reduces the standard deviation of prediction error by $31 \%$ compared with ridge regression. In particular, 58% reduction is observed for vehicles with large prediction error. Furthermore, the proposed method also reduces the average prediction error by $16 \%$ compared with ridge regression. This study contributes to improving operating efficiency by enhancing the accuracy of lifespan predictions for fuel-cell systems in industrial vehicles and advancing time-series prediction technology.