State of health estimation for lithium-ion batteries on few-shot learning

State of health (SOH) is a critical indicator for implementing detection, diagnostics and prognostics on lithium-ion batteries. However, considering the difficulty of data collection and additional cost for gathering comprehensive field data in practical application, only limited data can be available for model establishment. In order to handle this insufficient data scenario, a novel Bayesian deep neural network has been established and validated on few-shot learning. Moreover, from the perspective of feature extraction, degradation patterns extracted from temporal cyclic discharge profiles are utilized for reflecting degradation mode and operation state, while the Gramian angular field is proposed for data distribution learning and information enhancement. Different percentages of data are conducted on model training to compare the comprehensive performance on various features and state-of-art methods with the proposed method on few-shot learning. Ultimately, experimental results prove better adaptability, generalization and effectiveness of the proposed method on lithium-ion battery SOH estimation regardless of data size.

Automated summary

State of health (SOH) is a critical indicator for lithium-ion battery analysis, but limited data is available for accurate model establishment due to difficulties in data collection. To address this issue, a novel Bayesian deep neural network is proposed and validated on few-shot learning. Degradation patterns extracted from temporal cyclic discharge profiles are utilized for reflecting degradation mode and operation state, while the Gramian angular field is proposed for data distribution learning and information enhancement. Experimental results demonstrate the effectiveness of the proposed method for accurate estimation of lithium-ion battery SOH, irrespective of data size.