聚类分析
计算机科学
变量(数学)
传输(计算)
排名(信息检索)
电池(电)
电动汽车
编码器
电压
断层(地质)
电池组
模式识别(心理学)
电池电压
人工智能
数据挖掘
汽车工程
模拟
实时计算
旋转编码器
称重传感器
细胞
统计
标识
DOI:10.1038/s41598-026-79400-6
摘要
Identifying a faulty cell from battery-management records is difficult when operating trends affect all cells and different diagnostic indicators favour different candidates. Here we present PhyFuseCell, a physics-anchored ranking method for packs with variable cell counts. Both branches reference cell voltages to the pack median: one characterizes charge-residual magnitudes and trends, while the other learns normalized discharge patterns. A shared cell encoder and scorer, together with valid-cell masking, accommodate different pack sizes. The evidence scores are standardized within each pack and combined with equal weights. On CH-BatteryGen packs with 28 or 124 cells, the compact implementation achieved the highest mean Top-1 localization accuracy among the evaluated methods, at 98.0 ± 2.7% across five vehicle-held-out splits, 7.0 percentage points above a validation-tuned clustering baseline. Paired predictions showed how complementary evidence corrected competing cell rankings. In self-discharge transfer without target-group training, adding target-pack charge evidence increased mean accuracy from 61.0 to 83.8% for 28-to-124-cell transfer and from 66.3 to 94.7% in reverse. These results support organizing physical and temporal evidence at the cell level to improve fault localization across the evaluated pack sizes.
科研通智能强力驱动
Strongly Powered by AbleSci AI