Tesla Battery Pack Failure: The Warning Signs Are There 90 Days Out

Based on a research paper in preparation on early detection of battery pack failure from field charging data, one part of the Dr.EV battery model pipeline.

Battery packs that are about to fail do not behave like healthy ones, and the difference is visible in ordinary charging data three months before the pack is replaced. That is the finding, and the rest of this article is how we got there and what it is worth.

Our team has spent about two years on it. In the data a failure shows up as a pack replacement on the Tesla Service record, and that date is what everything here is measured against.

What goes wrong inside a pack

A Tesla pack is a large number of individual cells wired in series and parallel, and the pack is only ever as good as its weakest cell. In a new pack the cells are closely matched. As the pack ages they drift apart, which is normal and slow, and the battery management system corrects for it by balancing.

Packs that end up failing are the ones where a single cell or cell group stops keeping pace with the rest. Two mechanisms account for most of it. The first is a rise in internal resistance in one cell, from impedance growth at the electrode interface, which makes that cell swing further in voltage than its neighbours whenever current flows through the pack. The second is a soft internal short, which slowly bleeds charge from one cell so that it falls behind the others faster than balancing can bring it back.

What the research had to solve

We could not specify in advance what a pack heading for failure looks like. Any single reading from a car moves around for reasons that have nothing to do with pack health. The same vehicle charging on a cold morning at high state of charge behaves differently from the way it behaves warm and near empty, and those swings are larger than the effect we were looking for.

This is why the study uses a deep neural network rather than a threshold on a signal we picked ourselves. Instead of deciding in advance which reading matters and where to draw a limit on it, the model learns from a large body of real charging data what a healthy Tesla battery does across the conditions these cars actually encounter, and then measures how far a given session sits from that.

The problem was therefore separation more than detection. The model had to be precise enough that a genuinely abnormal pack stands out against ordinary variation instead of being lost in it. That model, and the study built around it, is what the paper describes.

We then went back through real Tesla charging data, including vehicles whose packs later failed, using only data recorded before the failure.

How we kept ourselves honest

It is easy to build something that looks clever on data where you already know the answer, so the study was arranged to make that difficult.

The model only ever studied healthy cars. It had never seen a failing pack before it was asked to pick one out. The vehicles it was measured on were vehicles it had never encountered at all. The alert threshold was chosen by looking at healthy cars, never by checking how many failures it would catch. And nothing recorded on or after the failure was used, only what the car produced while it was still being driven normally.

One thing we should say plainly. We chose the best-performing approach after seeing how several of them did, which is ordinary at this stage of research but does mean these numbers could come out lower when the method is tried on cars nobody has looked at yet.

Reading the charts

Every chart here plots the same quantity: the score our machine learning model assigns to one charging session. The model compares what a pack actually did against what a healthy pack would have done under the same conditions, at the same state of charge, the same current and the same outside temperature. The result is placed on a 0 to 100 scale, where around 50 is a typical healthy car and higher values mean the pack resembles a healthy one less and less.

A high score therefore means the session did not fit the model's idea of normal. A single one means very little, since any car can produce an unusual session. What matters is whether a pack keeps scoring high across many sessions.

1. Failing packs already look different three months out

The blue line is the healthy fleet, sitting at 50 across the whole period. The orange line is the packs that failed. In the 60 to 90 day band they are already at 84, they climb through the 90s, and they reach 98 in the final week before the failure. Ninety days is as far back as we looked and the separation is already established at that point, so this data cannot tell us when it begins.

The flat blue line carries as much weight as the orange one. Vehicle age and accumulated mileage are the obvious confounders, since a model that had simply learned to recognise old, heavily used cars would produce exactly the kind of separation we are claiming. If that were happening the healthy line would rise as well, because the healthy fleet contains plenty of old and heavily used cars. It stays flat.

2. Turning a signal into a warning costs earliness

An alert is the score crossing a threshold, and where that threshold sits determines everything else, because setting it low enough to catch subtle cases means sending warnings to healthy owners who do not need them.

At a threshold chosen to leave the large majority of healthy cars undisturbed, the model flagged just over half of the failing packs, with a typical first warning about 38 days before the failure. That figure needs context, because half sounds unimpressive on its own. The same threshold spends warnings on roughly 5 in 100 healthy cars, so a score carrying no information at all would land on about 5% of failing packs by chance. This one lands on more than 50%.

The bars fall as the requirement gets stricter. About a third of failing packs were flagged a month or more ahead, and roughly one in eight got two months or more. Being visible at three months is not the same as being warned at three months, and the gap between them is the price of keeping false alarms low.

Close to half of the failing packs were not flagged at this threshold at all, and some showed nothing unusual until the last few days.

3. The trade-off is unavoidable

Every warning system faces this curve and no amount of modelling removes it. Each point represents a different threshold. At the cautious end about 1 in 100 healthy cars ever sees a warning and roughly a fifth of failures are caught. At the sensitive end about two thirds of failures are caught and roughly 1 in 10 healthy cars receives a warning.

The orange point marks the setting used throughout this article, at around 5 in 100 healthy cars warned over a three month window and just over half of failing packs caught. We chose it on the view that an owner's tolerance for a false alarm is low but not zero, and that a warning at this level means the pack is worth inspecting rather than that it is failing.