
The Automotive domain is evolving fast. If 5 years ago software written in ECUs was simple and some might say even “dumb”, today we can see Machine Learning (ML) is here and will play an important role in how vehicles will behave in the future. Autonomous driving is not just a distant dream, but tomorrow’s reality, as intelligent vision, radar or lidar systems are becoming part of standard vehicle equipment. But applications for ML don’t stop there.
One of the biggest challenges in Electric Vehicles (EV) is the autonomy of the batteries, measured in both maximum charging cycles until full degradation and remaining kilometers until the next charge. How can this autonomy be extended by applying knowledge from the applied ML domain?
Data-driven prediction of the battery life cycle, such as the one presented in [R1], can help with the first challenge: determine the maximum charging cycles until full degradation. Data collected during numerous charging situations, including fast charging, varying cell temperature yield currents and voltages which are fed to mathematical models that can predict the battery degradation curve over time. Digital representations of the physical batteries can be used to speed up the prediction process. The digital copy and the original device shall start from the same set of parameters when they leave the production line. Over time, the digital copy will evolve based on data collected from the on-field physical battery. Running the ML models with the set of data mentioned above for several times (i.e. simulating same historical usage of the batteries in the same conditions repeatedly) will produce a prediction of the battery’s decay.
A second challenge involves determining the maximum distance until the next required charge based on driver habits. These include the way one accelerates, brakes, but also the driving routes, and daily routines e.g. number of stoplights to work, number of stops, or weekend drives. This profile can be learned (in the ML sense) and lead to automatic autonomous driving level IV [R2] adjustments of the usage of the batteries. We can imagine taking advantage of crowd-sourced driving experiences, i.e. millions of driving patterns and being notified e.g. “you can improve your car’s remaining battery life if you do X instead of Y”. Combining individual driving profiles with the shared general driving experience can optimize battery usage.
Going forward, the OEM can intervene and optimize battery usage by actively limiting the driver’s actions. Consider for example the case of a carsharing operator that may be interested in limiting the maximum acceleration of the vehicle, the recuperation (i.e. regenerative braking is an energy recovery mechanism that slows a vehicle or object by converting its kinetic energy into a form that can be either used immediately or stored until needed [R3] percentage or the numbers of fast charges in order to protect the vehicle’s batteries and increase their utilization period. All of this is possible if the operator has access to the driving profiles and behaviors resulting from applied ML models and taking constraints in the form of parametrized data sent over-the-air to the vehicles. Drivers can see this as a money-saving mode being activated (by default), compared to, say, a sports mode for which an extra fee would unlock the desired driving experience.
References:
[R1] “Data-driven prediction of battery cycle life before capacity degradation”, Kristen A. Severson, Peter M. Attia, Norman Jin, Nicholas Perkins, Benben Jiang, Zi Yang, Michael H. Chen, Muratahan Aykol, Patrick K. Herring, Dimitrios Fraggedakis, Martin Z. Bazant, Stephen J. Harris, William C. Chueh & Richard D. Braatz, Nature Energy volume 4, pages 383–391 (2019)
[R2] “Taxonomy and Definitions for Terms Related to Driving Automation Systems for On-Road Motor Vehicles”, SAE International, J3016_201806
[R3] “Regenerative brake”, Wikipedia