Exploiting domain knowledge to reduce data requirements for battery health monitoring
Publication Name
Energy Storage Materials
Abstract
Rechargeable batteries are becoming increasingly significant in decarbonising the world. For their widespread usage, to monitor and predict the battery health status has been essential. Although machine learning has the potential to tackle this issue, considerable degradation tests are required for model training, leading to prohibitive costs and labour. Here, we introduce a novel approach to constructing health monitoring models by fusing battery degradation knowledge with deep learning, using a substantially reduced amount of degradation data. We employ a lightweight and interpretable model to produce synthetic charging curves from highly limited realistic data. Subsequently, a transfer learning technique is implemented to train a convolutional neural network using both types of data and alleviate their gap. By employing only 8 realistic charging curves to develop the model, the method can precisely estimate the maximum and remaining capacities from 300 mV charging segments. The root mean square errors for these estimations are below 12.42 mAh. Additional 50 validation cases confirm that the proposed method can not only reduce the required degradation data but also shorten the input window length. Furthermore, it can be generalised and applied to different battery types under different operating conditions. This work highlights the promise of employing domain expertise to significantly decrease the amount of battery testing required for monitoring battery health.
Open Access Status
This publication is not available as open access
Volume
67
Article Number
103270
Funding Number
52207229
Funding Sponsor
National Natural Science Foundation of China