Online state-of-health estimation of lithium-ion battery based on relevance vector machine with dynamic integration

The lithium-ion battery's state of health (SOH) is one of the essential parameters of the battery management system. An accurate state of health estimation of the battery pack helps to improve the service life of the overall battery pack. Given the poor generalization ability of a single data-driven model during SOH online estimation and the lack of uncertainty expression ability in the estimation results, this paper proposes an online SOH estimation method of lithium-ion battery based on bat algorithm optimized relevance vector machine (BA-RVM) with dynamic integration. Firstly, we perform feature extraction and select equal voltage drop discharge time as an indirect health factor. Secondly, we establish the integration model, take the wavelet kernel relevance vector machine (RVM) as the sub-model, and use the bat algorithm (BA) to optimize its kernel parameters to improve the estimation accuracy of the sub-model. Then we use the online monitoring data to update the weights of the sub-models continuously and dynamically integrate the output of the sub-models to improve the accuracy of SOH online estimation further. Finally, the correctness and effectiveness of the method are verified based on battery data from NASA and compared with other data-driven methods. The experimental results show that compared with the method based on a single data-driven model, this method has higher accuracy and more vital generalization ability, and the estimation results have specific uncertainty expression ability. (c) 2022 Elsevier B.V. All rights reserved.

Automated summary

This paper proposes an online state of health estimation method for lithium-ion batteries based on bat algorithm optimized relevance vector machine (BA-RVM) with dynamic integration. The method improves estimation accuracy by performing feature extraction, establishing an integration model, continuously updating sub-model weights, and dynamically integrating sub-model outputs. It shows better generalization ability and uncertainty expression compared to other data-driven methods.