Limited labelled bearing-fault data are a practical challenge for EV gearbox condition monitoring, motivating diagnostic methods that can learn effectively from only a few examples. EV gearboxes also operate under varying speed and load conditions and may experience torque transients, which can alter vibration signatures and further challenge the generalization of diagnostic models trained with limited data. To investigate this problem in a controlled and reproducible setting, this study evaluates a Prototypical Network for few-shot bearing fault diagnosis using the publicly available CWRU vibration benchmark. A 4-way K-shot protocol is used with K = 1, 3, 5, and 10, and a lightweight 1D-CNN is employed to extract robust embeddings while preventing overfitting on small support sets. In the pooled-load benchmark evaluation, ProtoNet achieves 98.87% accuracy in the 1-shot case, 99.80% in the 3-shot case, 99.58% in the 5-shot case, and 99.78% in the 10-shot case, outperforming SVM, KNN, and a supervised 1D-CNN under the same comparison protocol. Additional leave-one-load-out tests across all three CWRU load permutations and K = 1, 3, 5, and 10 maintain mean accuracies above 99.80% over 1,000 independent test episodes per condition. These results demonstrate strong few-shot performance and robustness to the limited load shifts represented by the CWRU benchmark, providing encouraging methodological evidence for the intended EV gearbox monitoring context. Because CWRU does not reproduce the full speed, transient-torque, noise, and installation variability of an EV gearbox, hardware-specific EV validation remains an important next step.
