Leveraging AI for EV Performance Improvement

As commercial fleets adopt more electric vehicles, they need accurate state-of-health measurements and smart charging algorithms to ensure their EVs have minimum unscheduled downtime.

April 1, 2024

The electric vehicle (EV) industry is experiencing unprecedented growth, driven by technological advancements, environmental concerns, and government incentives. The total sales of EVs exceeded 10 million units in 2022 and are expected to reach 40 million units by 2030.¹ These significant growth rates are expected in all vehicle segments, including passenger EVs, light commercial vehicles, school buses, and heavy-duty trucks. At the same time, several challenges are becoming apparent in the adoption of EVs for individuals and businesses.

Due to the restricted availability of raw materials and environmental concerns, manufacturers must significantly extend battery life for first use in automotive and secondary use in other industries. There are discrepancies between the EV range measured by government-regulated standards and those observed in real life, leading to uncertainty about EV performance.² To maintain the residual values of EVs, the health of batteries must be accurately estimated since the battery can cost a major fraction of the total cost of the vehicle.³

As commercial fleets adopt more electric vehicles, they need accurate state-of-health measurements and smart charging algorithms to ensure their EVs have minimum unscheduled downtime.⁴ To solve these problems effectively, we need comprehensive data collection, capable computing infrastructure, and intelligent algorithms.

Technology Trends

The good news is that technology trends in automotive are heading in the right direction.

First, let’s look at data collection. Semiconductor manufacturers like NXP and Texas Instruments are building high-accuracy and low-latency sensors to measure battery parameters, such as voltage, current, and temperature.⁵ We need such high-resolution and high-frequency data to estimate battery parameters accurately since slight differences in voltage measurements can indicate significant differences in parameters like the State of Charge (SoC).⁶, ⁷

Next, there is an encouraging trend towards zonal architectures with high-performance computing engines in charge of each zone in the car.⁸ This trend started with high-performance chips like the Qualcomm Snapdragon in the infotainment/cockpit domain and the Nvidia Drive Orin in the ADAS (Advanced Driver Assistance Services) domain.⁹ Over time, we see higher computation availability in domains like the battery management system (BMS).

The more powerful chips will allow for rich computation on the edge to process data related to the BMS. These computing chips support modern software toolkits like TensorFlow-lite, enabling easy development of AI architectures, and middleware like Blackberry IVY that can absorb all the data available within the CAN network and feed it into AI applications.¹⁰

Further, the telematics units in vehicles have seen significant upgrades with the capabilities to transmit and receive data at 5G rates. Standard software protocols like MQTT allow for easy development of software transmission and reception capabilities on edge devices.

This infrastructure allows for data to be transmitted from vehicles to the cloud for further processing and, on the return path, for software in the cloud to control and improve the performance of vehicle components.

Finally, cloud computing has become pervasive and easy to implement in multiple domains. It is relatively easy now to handle computational tasks in the cloud with deadlines of a few milliseconds. This allows retraining algorithms for battery models to run efficiently and at minimal costs.

Potential Solutions

With these technology trends in progress, the time is ripe for applying advanced AI/ML software to improve overall battery performance in the car. All the major technology components are in place, and it’s a matter of integrating them into the appropriate products that deliver value to battery manufacturers, OEMs, fleet managers, and end customers.

Let us first take the example of AI for battery range estimation. Electra Vehicles has demonstrated its AI-based product for range estimation called Range Expert. Range Expert runs as an AI-based application on Blackberry IVY in a Qualcomm Snapdragon chip.¹¹

This algorithm uses TensorFlow-lite running in IVY and consumes a tiny fraction of the Snapdragon to deliver accurate range estimates. These range estimates can achieve up to 1 percent estimation error for 30-minute or shorter trips and up to 2 percent error for 1-hour or shorter trips.

Effectively, such advanced range estimation techniques can make range anxiety a problem of the past. This is just the tip of the iceberg. Over time, we will see more advanced AI-based algorithms applied to battery-related problems run in the infotainment units.

In other applications, they will run on Nvidia GPUs to recommend battery-optimal routes automatically chosen by the ADAS systems. End users will reap the benefits of this sophisticated processing with fewer recharging stops and longer battery life.

As a second example, we consider SoC estimation. The traditional method of Extended Kalman Filter (EKF) has been used extensively for this purpose. EKFs demonstrate high accuracy at the beginning of the battery life, but the performance degrades over time. Neural networks with adaptation based on collected data can solve this problem.

Figure 1. Electra’s Adaptive Cell Modeling System (ACMS) is 2x more accurate than EKF after nine weeks of accelerated battery aging. (Image: Electra)

As shown in Figure 1, Electra Vehicles has shown a 2x performance improvement compared to EKF using such an approach in a bench setup. Further, the computational workloads required for such inference engines can be optimized for embedded targets. For example, the algorithm described in (BlackBerry, 2023b) takes only 1.2 ms on an ARM Cortex-M4 at a clock rate of 160 MHz (a typical MCU configuration used in BMS systems).

Finally, we turn to next-generation technologies like distributed ledgers for battery applications. The European Union has taken the lead in building a “Battery Passport.”¹² From February 2027, every EV sold in Europe will need a digital battery passport for its battery.¹³

The battery passport must contain comprehensive historical information on the State of Health (SoH) of the battery.¹⁴ This will need information relayed from the embedded BMS to the cloud, computation of an accurate SoH in the cloud, and communication of that SoH to a distributed data storage system that implements the battery passport.

For example, Electra Vehicles has a cloud-based Fleet Analytics product that provides accurate estimates of SoH and the resulting residual value of electric vehicle batteries.

Future of Electrification

Artificial intelligence and machine learning applications in automotive have been primarily restricted to ADAS systems. However, the technologies driving these applications can now be used for EV battery performance and monitoring. There are some key engineering challenges to be solved. For instance, all data must be handled with the highest privacy and security standards.

Further, the increased performance from the AI/ML-based solutions must justify the increased operational costs. However, from the demands in the industry and the march of technology, it is clear that these challenges can be overcome and that AI/ML-based solutions will play a major role in the future of electrification.

This article was written by Dilip Warrier, Ph.D., VP of Engineering at Electra Vehicles (Boston, MA). For more information, visit here .