(632b) Improved State-of-Health Estimation for Batteries with Domain-Adaptive Deep Learning

Lithium-ion batteries are presently considered key components across various fields, offering vital power sources for everyday devices like smartphones and laptops, and extending their utility to renewable energy infrastructures, healthcare apparatus, etc., due to their exceptional energy storage capacity and rechargeability [1-2]. Monitoring the state-of-health (SOH) of these batteries through accurate capacity estimation is crucial for predicting battery remaining lifetime and undertaking corrective actions before unexpected failures. Deep learning models offer advanced capabilities for SOH estimation in batteries by learning complex patterns from historical data to predict future performance with high accuracy [3]. However, the diverse operational conditions and degradation behaviors of these batteries lead to largely distinct distributions in the corresponding data [4]. As a result, DL-based SOH prediction models trained with a specific operating data (source domain) often fail to generalize to new, unseen data (in the target domain). To bridge this gap, domain adaptation has been adopted in the machine learning community as a tool to ensure that predictive models remain reliable across various domains [5-6].

In this study, we address the above domain shift issue for battery SOH estimation with a domain adaptation-based DL technique based on maximum mean discrepancy (MMD). As a kernel-based metric, the MMD measures the distance between the mean embeddings of two different data distributions in a high-dimensional feature space (via the kernel mapping) [5]. We assume that the source domain data are supervised with labels (capacity values) for SOH estimation whereas the target domain data are unsupervised without labels, which is common in practice due to the high cost of acquiring labels [4,7]. The presented domain-adaptation DL framework consists of dynamic feature extraction layers based on long short-term memory (LSTM) networks, followed by fully-connected (FC) layers for capacity estimation. MMD is applied to the outputs of FC layers to measure the difference between FC features acquired from source and target domains. The total loss for the optimization includes both capacity prediction (regression) loss for source data and the MMD loss between two domains. Minimizing the MMD loss can align the source-domain features with target-domain features, making the source and target domains more statistically similar, and thus enhancing the generalizability of trained models to the target domain [4, 7].

In the case study, we considered the NASA battery dataset (with labels) as the source domain and the Michigan battery dataset (unlabeled) as the target domain [8, 9]. These two batteries had their aging test experiments at two different conditions (room temperature for NASA and -5^oC for Michigan). To evaluate the robustness of our domain-adaption DL method, we tested our model with some labeled data from the Michigan dataset. We compared our result to a model completely trained on the source domain without any domain adaptation. Currently, our results give a root mean square error (RMSE) value of 0.188 for the model with domain adaptation and 0.343 for the case without domain adaptation This represents a significant improvement in model adaptability with domain adaptation based on MMD. Thus, our work shows that domain adaptation can serve as an effective means of carrying out SOH estimation for batteries under different working conditions.

References:

[1] Grey, Clare P., and David S. Hall. "Prospects for lithium-ion batteries and beyond—a 2030 vision." Nature communications 11, no. 1 (2020): 1-4.

[2] Ralls, Alessandro M., Kaitlin Leong, Jennifer Clayton, Phillip Fuelling, Cody Mercer, Vincent Navarro, and Pradeep L. Menezes. "The Role of Lithium-Ion Batteries in the Growing Trend of Electric Vehicles." Materials 16, no. 17 (2023): 6063.

[3] Kaur, Kirandeep, Akhil Garg, Xujian Cui, Surinder Singh, and Bijaya Ketan Panigrahi. "Deep learning networks for capacity estimation for monitoring SOH of Li‐ion batteries for electric vehicles." International Journal of Energy Research 45, no. 2 (2021): 3113-3128.

[4] Ye, Zhuang, Jianbo Yu, and Lei Mao. "Multisource domain adaption for health degradation monitoring of lithium-ion batteries." IEEE Transactions on Transportation Electrification 7, no. 4 (2021): 2279-2292.

[5] Long, Mingsheng, Yue Cao, Jianmin Wang, and Michael Jordan. "Learning transferable features with deep adaptation networks." In International conference on machine learning, PMLR (2015): 97-105.

[6] Liu, Xiaofeng, Chaehwa Yoo, Fangxu Xing, Hyejin Oh, Georges El Fakhri, Je-Won Kang, and Jonghye Woo. "Deep unsupervised domain adaptation: A review of recent advances and perspectives." APSIPA Transactions on Signal and Information Processing 11, no. 1 (2022).

[7] Han, Te, Zhe Wang, and Huixing Meng. "End-to-end capacity estimation of Lithium-ion batteries with an enhanced long short-term memory network considering domain adaptation." Journal of Power Sources 520 (2022): 230823.

[8] Saha, Bhaskar, and Kai Goebel. "Model adaptation for prognostics in a particle filtering framework." International Journal of Prognostics and Health Management Volume 2 (color) 61 (2011).

[9] Mohtat, Peyman, J. Siegel, A. Stefanopoulou, and Suhak Lee. "Uofm pouch cell voltage and expansion cyclic aging dataset [data set]." University of Michigan—Deep Blue Data (2021).