State Estimation for Battery Populations by Holistic Analysis of Ultrasonic Waveforms

Battery inspection using ultrasonic waves has gained popularity since 2015 [1] as a non-electrochemical method to estimate battery states. The propagation of sound through the internal structure of a battery is influenced by the mechanical properties and geometry of its constituent layers, which are known to vary during cycling and as the battery ages. In particular, alterations in the stiffness, density, thickness and porosity of the layers affect the speed of sound, its attenuation and length-of-travel. As a result, ultrasonic waveforms recorded during battery operation can be correlated to its internal states, and the accuracy of this mapping is subject to the modelling method used. Most attempts at modelling the acoustic signatures of batteries simplify the task by isolating specific waveform peaks, most commonly the final peak which is an ‘echo’ originating at the posterior side of the cell [2]–[5]. However, different battery layers will influence different waveform portions disproportionately, and such localisations in the analysis could lead to imprecision. More importantly, such localisations do not produce consistent trends across different cells, as shown in the accompanying figure. The figure depicts results from acoustic experiments conducted on seven commercial pouch cells (cells A to G). The cells were of the same type (product id) and chemistry, comprising a graphite negative electrode and a lithium cobalt oxide (LCO) positive electrode. The cells were tested for a total of 45 cycles using CC-CV protocols, out of which 25 cycles were performed at a rate of 1C and are the ones plotted. The remaining 20 cycles were performed at different rates, at intermediate times, and are not shown. The second and last acoustic peaks were isolated and their amplitude and Time-of-Flight (ToF) are plotted for each cell against battery capacity (Q). It is evident that the cell-to-cell variation of these localised acoustic characteristics is large, to the extent that a model relying on these features alone will be unable to generalise to the entire population. It should be noted that the temperature variation among experiments was 8 °C (18.8 - 26.8 °C) which does not explain the cell-to-cell variation seen. Alternative analysis methods that can utilise waveforms in their entirety were proposed in previous work [6]. Artificial neural networks were used to estimate battery state-of-charge (SoC) using multiple features from the time and frequency domains. In the present study we explore extensions of this machine learning approach to exploit feature spaces of even higher dimensions, and we assess the generalisation performance of these state estimators compared to methods based on peak selection. [1] A. G. Hsieh et al. , “Electrochemical-acoustic time of flight: in operando correlation of physical dynamics with battery charge and health,” Energy Environ. Sci. , vol. 8, no. 5, pp. 1569–1577, May 2015, doi: 10.1039/C5EE00111K. [2] P. Ladpli, F. Kopsaftopoulos, and F. K. Chang, “Estimating state of charge and health of lithium-ion batteries with guided waves using built-in piezoelectric sensors/actuators,” J. Power Sources , vol. 384, pp. 342–354, Apr. 2018, doi: 10.1016/J.JPOWSOUR.2018.02.056. [3] L. Gold et al. , “Probing lithium-ion batteries’ state-of-charge using ultrasonic transmission – Concept and laboratory testing,” J. Power Sources , vol. 343, pp. 536–544, Mar. 2017, doi: 10.1016/J.JPOWSOUR.2017.01.090. [4] G. Davies et al. , “State of Charge and State of Health Estimation Using Electrochemical Acoustic Time of Flight Analysis,” J. Electrochem. Soc. , vol. 164, no. 12, p. A2746, Sep. 2017, doi: 10.1149/2.1411712JES. [5] R. E. Owen et al. , “Operando Ultrasonic Monitoring of Lithium-Ion Battery Temperature and Behaviour at Different Cycling Rates and under Drive Cycle Conditions,” J. Electrochem. Soc. , vol. 169, no. 4, p. 40563, Apr. 2022, doi: 10.1149/1945-7111/ac6833. [6] E. Galiounas, T. G. Tranter, R. E. Owen, J. B. Robinson, P. R. Shearing, and D. J. L. Brett, “Battery state-of-charge estimation using machine learning analysis of ultrasonic signatures,” Energy AI , vol. 10, p. 100188, Nov. 2022, doi: 10.1016/J.EGYAI.2022.100188. Figure 1