Uneven internal SOC distribution estimation of lithium-ion batteries using ultrasonic transmission signals: A new data screening technique and an improved deep residual network

Abstract

Ultrasonic for state of charge (SOC) estimation of lithium-ion batteries has the advantages of non-destructive and real-time. The existing methods mainly depend on single-site detection, which is based on the assumption of uniform SOC distribution. However, the uneven SOC distribution existing inside the cell will cause rapid degradation of local performance, thereby bringing safety risks. Therefore, a novel method combining multi-site detection signals for the uneven internal SOC distribution estimation has been proposed, including Gaussian process regression-active learning (GPR-AL) and deep residual-pooling extreme learning machine (DR-PELM). Firstly, a focused ultrasonic beam is adopted to scan the cell. The preferred sites with lower uncertainty and their signal amplitude of ultrasonic waveform are extracted by GPR-AL. Then, DR-PELM has been established to learn the relationship between ultrasound signal features and SOC, which can reduce the impact of redundant information and noise. Finally, the accuracy of method has been verified through several case studies and destructive tests of lithium-ion detection. The results show that the mean error of general SOC estimation is 2.88 %, and the uneven SOC distribution estimation error is 0.37 %. Thus, the proposed method present good accuracy by integrating multiple selection sites with lower uncertainty and optimizing the network structure.