Enhancing Battery Safety in Electric Vehicles Through Predictive Analytics

A BMS is a device composed of both hardware and software. It is the BMS’s job to ensure at least three main functions: the first and most important being safety, the second is to ensure good battery lifetime and the third function is to ensure performance.

In terms of safety, what we really want is to avoid catastrophic events such as fires or explosions. For lifetime, we are looking to ensure that the battery will last as long as it can – both from a warranty point of view and also for the satisfaction of the customer. On performance, we need to make sure that the battery can deliver the energy and the power that the electric motor needs.

For example, we expect the BMS to provide an estimation of State of Charge (SOC), an estimation of State of Health and communicate these data to the Vehicle Controller Unit (VCU). In a nutshell, a BMS is a combination of all three of these aspects, ensuring them with a variety of hardware and software elements.

In order to avoid critical events, you need continuous monitoring. For that, we need to be monitoring the most relevant parameters of the battery pack, which are usually the voltage, current and temperature. Because we are addressing lithium-ion batteries, we need to not only monitor the global battery pack, but also the single-cell voltage and temperatures to make sure that we are operating in a safe environment.

Safety can also be improved through monitoring what we call “overcharging”. Especially when the vehicle is coming into a charging station, you want to make sure that the charging station is communicating properly with the BMS to adapt the charging profile of the battery pack based on SOC, SOH and temperature.

Improving safety is also about avoiding thermal runaway. This is one of the most critical events that can lead to fires and explosions if there is no intervention. Thermal runaway will happen when the temperature of the cells reaches unwanted high values until a point where it can not be controlled anymore, even if the power supply is cut.

The BMS’s core competency is really to monitor all these aspects to ensure better safety and efficiency.

That is a very interesting question, because if you were to rewind just a little bit in time and go into the literature, you would see that this concept of a battery’s “state of health” (SoH) has covered many different notions and definitions.

Essentially, we can think of a battery’s SoH as being like that of a human being’s; sometimes you are ill, but you can still heal and recover, sometimes you are ill, but it is an irreversible, terminal disease.

The most difficult thing for SoH is to define what it is. Back in the past, we had different notions about SoH as a function of energy loss versus SoH as a function of power loss. Both are critical for a battery and a vehicle – if you are losing energy, that means you have less range in your EV, while if you are losing power, then you have less acceleration and torque available for the electric motor.

Today, the industry has coalesced around SoH as an energy-oriented term.

One of the biggest challenges is to estimate SoH without going into the cell and discharging it to control the real values of internal resistance, capacity and so on. A non-invasive solution would be great, but this is not easy when you have a product embedded within a vehicle.

Unfortunately for us, SoH is also very non-linear, which means it can be difficult to predict SoH evolution because it depends on many different parameters. For example, what temperatures has the battery seen during its lifetime? What is the pattern of its charging sessions? Was the battery potentially misused? SoH is always a very tough parameter to access, but we are currently progressing in this field thanks to a range of new, innovative techniques.

If you look at BMS functions, the algorithms that are feeding into the BMS are usually physics-based, using something called equivalent circuit modelling. Essentially, this attempts to consider the battery as a purely electric device. We can refine this model using an electrochemical model, which brings much more detail about the structure of the battery and the electrodes. But it is complicated to bring this into an embedded system like a BMS.

You also have some other techniques, such as the extended Kalman filter, which uses the electric model and an electrothermal model as a basis for extrapolating out some values related to the battery pack, such as its state of charge or SoH. Depending on what we are measuring through the BMS, we can calculate the error between these predictions in the short term and what is measured. Then we re-inject this error into the electric model, effectively creating a kind of self-learning model.

That is the basics of BMS today. AI is usually used to complement this physics-based model, bringing in additional data and relying on neural networks to improve the accuracy, reliability and fidelity of the model. In effect, what you have here is a blend between the physics and the AI.

In certain cases, you can observe some players in the space who are completely omitting the physics and choosing to work with just AI and neural networks – training this network with huge amounts of data on voltage, current, temperature and so on as its inputs. Then they just use this AI model as a BMS algorithm – and it does give some pretty good results.  It shall be noted that collecting all these data and training the neural network can be extremely time-consuming.

In summation, you have these two approaches with AI: either using it to complement your physics-based model or solely relying on AI as your algorithm. 

Predictive analytics is something that is relevant for BMS because, beyond just ensuring battery safety and performance, you also allow the final customer or original equipment manufacturer to be able to predict when the battery is going to die.

When I talk about a battery dying, I am referring to this notion of what we call the RUL – remaining useful life. What we are trying to do here is to predict how long a battery will last if it continues to be used in the same conditions it has been used previously. In the automotive industry, the criterion is generally that a battery is considered “dead” when its SoH falls under the 80% mark (or when internal resistance is doubled). Once this 80% mark is reached, the battery can still be very useful and be given a second life in other applications, but it is just no longer optimal for EVs.

What is interesting is when you have a “middle-aged” battery that’s at 90% and you want to know when it will hit this 80% value, perhaps for warranty purposes or for the general customer satisfaction of knowing that they are not going to be stuck in the middle of nowhere on the road. In short, predictive analytics is used to make RUL relevant in the long run.

There is another side to prediction, which is more related to safety. From a regulation standpoint, you have some rules, especially relevant in China, stating that you must be able to detect thermal runaway within five minutes in order to leave enough time for the occupants of that vehicle to get out of the car and away from danger. That’s an issue of detection. But there are also some new trends emerging that are helping to predict thermal runaway much longer in advance than five minutes. That’s another example of predictive analysis being used, here to provide warning of failures.

In the future, I expect that the BMS will become more and more intelligent. For that, we will want to be using all the techniques that we have discussed here – AI, machine learning. These predictive analytics tools will be used to ensure that the battery pack is going to be able to recharge properly.

Something that I also think we will also see in the future are “digital twins” being merged with Cloud BMS – a BMS that runs in the cloud – where you have potentially almost no limits on computation. Once the computation is done in the cloud, this information can be sent back to the embedded BMS so that you have the most relevant up-to-date data in the vehicle.

I think there will also be more progress regarding the detection and prediction of thermal runaway because, as I said, we are talking about safety here and this is non-negotiable.

I think BMS should also be more agnostic to battery chemistry. As we see today, we have a lot of lithium-ion chemistries – nickel manganese cobalt (NMC), lithium iron phosphate (LFP), lithium manganese oxide (LMO) and others – but we also see some other chemistries starting to pop up, such as sodium-ion batteries, solid state batteries and lithium-sulfur batteries. I think the industry will try to make the BMS as agnostic as it can be to chemistry, because we don’t know what is going to win out in terms of being the dominant chemistry in the future.

Perhaps one of the challenges I foresee is to do with connecting the BMS to the cloud. In Cloud BMS, you are exchanging data and so in the European Union, we must obey General Data Protection Regulation rules. There will need to be BMS with cybersecurity features that are built-in to the embedded BMS.