This paper describes a means to predict the internal structure of a lithium-ion battery from the response of an ultrasonic pulse, using a genetic algorithm. Lithium-ion batteries are sealed components and the internal states of the cell such as charge, health, and presence of structural defects are difficult to measure. Ultrasonic inspection of lithium-ion batteries is a recent and growing area of research. Reflected and transmitted ultrasound pulses are proposed as a non-invasive means of gaining insights into the internal structure and changes within the closed body of a cell. However, the multiple layers present in a lithium-ion cell are problematic when attempting to interpret waveforms as many internal reflections superimpose. Attributing specific features of a cell to wave characteristics is challenging. In this work a genetic algorithm has been developed as a means to reverse engineer a single ultrasound wave response to predict the internal layered structure of a lithium-ion battery cell. A first randomised guess at the layered structure is made. A numerical wave propagation model is used to predict the ultrasound waveform associated with that structure. This waveform is then compared with a measured or reference waveform to establish its fitness. The layered structure is generationally mutated until the predicted waveforms converge on the reference signal. As this occurs the predicted layered body reveals insights into the cell structure under inspection. Initially, the algorithm was tested against an idealised model battery and its predicted waveform, giving a model-model verification. Further, experimental ultrasonic reflection signals were captured from small capacity lithium-ion cells. Estimations of layer structure predicted by the model were compared with CT-scans of the cells to assess performance. The genetic algorithm was found to be effective in converging the predicted wave response to the reference signal and creating accurate battery structures. It was shown that only part of the waveform was required to generate accurate predictions, which is helpful in avoiding parts of the signal contaminated by near field transducer effects. It was demonstrated that the genetic algorithm can predict material wave speed to 40–1100 m/s (3–29 %) accuracy when battery layer geometry is provided; and thicknesses to within approximately 0.2–7.5 μm (1–13 %) when material properties are provided. Providing the genetic algorithm with parameter constraints; either the layer topology and/or the material properties, substantially improved predictions to estimate the wave speeds on average to approximately ±50 m/s (3–4 %) and the layer thicknesses ±5 μm (7–8 %). This raises the possibility of the use of this approach to predict state of charge when the battery construction is known, or the presence of internal defects and damage to a known battery material composition.
CORE (COnnecting REpositories)
CORE (COnnecting REpositories)