可靠性工程
可靠性(半导体)
计算机科学
电池(电)
可靠性理论
扩散
数学模型
工程类
概率密度函数
汽车工程
材料科学
电子工程
模型验证
控制理论(社会学)
功能(生物学)
瞬态分析
估计理论
电流(流体)
作者
Zhiyi Xu,Armin J. Köhler,Martin Dazer
标识
DOI:10.1109/tr.2026.3749192
摘要
Reliable battery diagnosis is important for maintaining Safety-Related Availability in fail-operational powernets. SOTIF-oriented validation of battery diagnostic functions requires extensive physical testing over diverse battery states, aging conditions, temperatures, and load profiles. This work investigates conditional generative modeling for battery voltage response generation using measurements from heterogeneous field-aged automotive low-voltage batteries. We first compare several representative generative models for this task. Among the evaluated models, the Diffusion Transformer (DiT) achieves the best overall distributional performance, with a sequence energy score of 0.168 V, precision of 0.996, and recall of 0.903. We then investigate whether transfer learning can reduce the data requirement by the selected DiT model. With only 10% of the target data, the source-pretrained and fine-tuned model already outperforms a target-only model trained with 25% of the target data in sequence energy score and the model trained with 50% in recall. This transfer benefit is also preserved in the downstream diagnostic evaluation using a fixed battery diagnostic function. Finally, we examine whether physics guidance further improves transfer learning. In contrast to the direct DiT model, the Physics-Residual model exhibits negative transfer. Further analysis shows that although fine-tuning reduces the target-domain MAE of the physics branch from 0.358 V to 0.173 V, adaptation of the physics branch increases the residual complexity. This indicates that physics guidance and transfer learning are not necessarily complementary. A controlled simulation study further demonstrates how a well-trained generative model can support hazardous diagnostic scenario exploration and subsequent SOTIF-oriented analysis. Overall, the results demonstrate the potential of transferable generative models to reduce project-specific battery testing and highlight the need for careful use of physics guidance when combining with transfer learning.
科研通智能强力驱动
Strongly Powered by AbleSci AI