Battery power supply systems are vital for the uninterrupted operation of industrial products, necessitating robust anomaly detection to maintain reliability. Traditional detection methods often struggle with feature extraction from long, non-stationary data and with the highly imbalanced nature of battery monitoring time series. To address these challenges, we propose CRLBAD (Clusterwise Representation Learning for Battery Anomaly Detection), an unsupervised anomaly detection framework tailored for imbalanced battery data. CRLBAD automatically extracts semantic features through representation learning, then employs DBI-optimized hierarchical clustering to partition the sample space, thereby mitigating data imbalance and enabling more effective outlier identification. Within each cluster, state-of-the-art unsupervised anomaly detection algorithms are used to detect abnormal batteries. Experimental results on a real-world dataset demonstrate that CRLBAD effectively identifies all anomalies and achieves superior AUC-ROC and AUC-PR scores, confirming its efficacy.