Fig. 3: Evaluating prediction accuracy and EV battery cost.

a The average ROC curves for the five algorithms. The solid curves indicate the average values out of five cross validation runs, and the shaded regions indicate the standard deviations of the trials. Our proposed procedure (orange) significantly outperforms other deep learning baselines (green, blue, yellow) and the standard variation evaluation method (gray). b The five-fold average AUROC of our proposed DyAD with different weights of the auxiliary training losses (KL regularization and mileage label loss). The variation of the average AUROC indicates that the auxiliary training losses can nontrivially improve our model performance. The cost ranges can be found in Methods. c The sum of direct fault cost and inspection cost for EV batteries based on the statistics (battery fault rate, battery fault cost and vehicle inspection cost) we collected from an EV platform. The horizontal axis indicates the true positive rate achieved by each model for the corresponding cost value. d The minimum cost achieved for each algorithm by optimizing the total cost against the true positive rate. The confidence range is evaluated from our estimation of the vehicle fault rate (from 0.038% to 0.075%).