The rapid deployment of battery swapping stations necessitates scalable and reliable fault diagnosis, yet massive, sparse operational data and scarce labeled samples make this challenging. Here, we report a rapid unsupervised learning framework for realistic  multi-fault diagnosis in million-scale battery fleets. Our approach employs a double-layer mechanism. First, we rapidly screen for abnormal devices by extracting features from voltage-envelope sequences. Subsequently, we pinpoint faulty cells and types using an enhanced two-stage unsupervised clustering combined with rule-based fault tracing. The framework is validated on a production dataset of over 128,000 devices, achieving 97.33% device-layer and 99.66% cell-layer accuracy. Laboratory tests on recalled batteries further confirm the detection of low-capacity and micro-short-circuit faults. These results demonstrate scalability and robustness under sparse-data conditions, enabling reliable operations for large-scale energy storage systems.