<p>In real-world electric vehicle (EV) operation, battery fault characteristics are strongly affected by operating conditions, while labeled fault samples are often scarce due to the low occurrence rate of battery faults. This paper proposes an integrated framework named FiLM-DPNet, a FiLM-modulated dual-branch prototypical network built upon prototypical meta-learning, for cross-condition few-shot fault detection. Specifically, each meta-task is constructed with a support set and a query set under the episodic few-shot learning paradigm. Fault-sensitive main features and condition-related auxiliary features are first extracted through two parallel temporal encoding branches. The auxiliary branch then generates feature-wise modulation parameters to adaptively recalibrate the fault-sensitive representation through Feature-wise Linear Modulation (FiLM). Based on the fused support features, an adaptive weighted multi-subprototype mechanism is developed to construct multiple representative subprototypes for each class, describing intra-class structural variations caused by operating-condition differences. Finally, query samples are classified according to the scaled cosine similarities between query embeddings and the constructed class subprototypes. Experiments on real-world EV operational data across six typical operating conditions show that, under the 5-shot setting, FiLM-DPNet achieves an average accuracy of 91.93% in the leave-one-condition-out evaluation, outperforming representative baselines in most tasks. In addition, FiLM-DPNet achieves an average accuracy of 89.62% in the leave-one-vehicle-out evaluation, providing a stricter assessment of unseen-vehicle generalization within the current data source. These results suggest that the proposed method provides a potential solution for EV battery fault detection with limited labeled fault data.</p>