Researchers from Technical University of Braunschweig found the method for early identification of the starting destruction of lithium-ion batteries by simply listening to the sounds they produce during operation. Apparently, micro-reactions are constantly going on inside batteries, and they are accompanied by hardly distinguishable acoustic pulses. These sounds are too weak to be heard, but quick-response piezo-sensor is capable of capturing them and transforming them into electric signal.
Each pulse carries information about the processes going on inside the battery. If electrolyte fluid starts degrading and emanates gas, the sound is soft and extended. If a graphite particle cracks in anode, the signal becomes short and sharp – like a micro-explosion. The team from Braunschweig decided to systematize these sounds, so that it becomes possible to judge about the battery’s condition and to forecast its wear-and-tear.
To do that, the researchers reproduced two types of degradation. In the first case, the battery was purposefully charged above the safe level causing violent gassing. In the second case, they used special solvent capable of breaking the graphite electrode. During these trials, the piezo-sensor registered thousands of acoustic events, and the computer processed each signal calculating twenty physical and statistical parameters – amplitude, energy, length, frequency spectrum and wave shape, etc. The researchers set these parameters in advance to receive quantitative description of what was going on inside the cell.
Then they turned to machine learning. At first, they used Isolation Forest algorithm, which learned to distinguish the sounds emerging during gas emanation, and the signals associated with destruction of materials. The obtained data were mapped out and used for training Random Forest classifier capable to distinguish which type of degradation cause a specific acoustic pulse. This simulation demonstrated about 90% accuracy successfully classifying new signals without human involvement.
The researchers checked the simulation performance using the full-sized battery of NCM-Graphite type after it had undergone 100 charge and discharge cycles. The simulation correctly identified more than half of all the acoustic pulses with over 75% level of confidence. The hyperactivity periods coincided with the phases of effective chemical and mechanical processes – protective cover formation, gas emanation and cyclic expansion of graphite.
Hence, the researchers proved that we can “hear” the battery without any invasive techniques and use the sound profile to judge about the reactions going on inside.
In future, such systems may become a part of standard modules of batteries control, automatically recording early signs of degrading or overheating.
