Reliable anomaly detection in electric vehicle (EV) batteries must be achieved without compromising user privacy, a challenge that grows with the increasing scale and diversity of EV fleets. This work introduces a hybrid federated split learning framework designed to support collaborative model training while keeping sensitive data local. Unlike previous approaches that either compromise privacy through centralized data collection or impose excessive computational burdens on resource-constrained devices, the proposed approach uniquely balances these competing concerns through an adaptive neural network splitting strategy. Based on the processing capacities of EVs and edge infrastructure, the framework dynamically allocates computational workload between them according to device capabilities. It employs a multi-layered privacy protection mechanism combining secure aggregation for model updates and homomorphic encryption to protect intermediate representations. Theoretical analysis shows that this approach reduces computational demands on resource-constrained EVs while offering strong resistance to inference attacks. Experiments on the EVBattery dataset demonstrate good anomaly detection performance and improvement over baselines with modest computational overhead. Its hierarchical architecture and adaptive splitting ensure resilience and efficiency across heterogeneous EV fleets with varying hardware capabilities.