Battery health from flight logs alone.

What state of health means for a drone battery

Every charge and discharge cycle degrades a lithium battery a little. State of health tracks that degradation as remaining capacity: a 2,200 mAh pack that can now hold 1,980 mAh is at 90 percent SoH. Industry practice, and the paper, treat 80 percent as end-of-life, the point at which a pack should leave the rotation before a cell fails outright. Internal resistance is the other health indicator, but neither it nor remaining capacity can be measured directly in flight.

For a fleet that makes SoH the number everything else depends on: how long a mission can safely run, when a pack should be retired, and whether the spares budget is going to packs that still had life in them. Predict it accurately and battery maintenance becomes a scheduling decision. Get it wrong in the optimistic direction and the failure mode is an aircraft coming down mid-mission.

Why drone batteries are harder to model than EV batteries

Machine-learning SoH prediction is a mature topic for electric vehicles. UAVs had been largely overlooked, and not for lack of interest. The constraints are different in five ways:

• Data scarcity. Training data comes from run-to-failure life-cycle tests, and for drone packs almost none had been published. Each pack has to be flown repeatedly until it dies.
• Variety. A single airframe accepts packs of many brands, sizes, capacities, and voltages, and an operator typically owns several. One model has to cope with all of them.
• No room for sensors. Temperature is one of the most informative signals of battery health, but a temperature sensor is extra weight, so the paper deliberately works without it. The inputs are limited to what the flight controller already records.
• Human and environmental noise. Pilot habits, payload, wind, and weather change how a pack is loaded from one flight to the next, in ways an EV drivetrain never experiences.
• Expensive ground truth. Remaining capacity can only be measured by fully discharging the pack, which itself damages it, so a label cannot be collected after every flight.

Large time-series models such as transformers, the state of the art for battery forecasting when data is plentiful, overfit badly on datasets this small. The research needed a method that could learn from a few hundred flights rather than a few hundred thousand.

How the data was collected: 631 flight tests

Rather than cycling packs on a bench, the team discharged them the way an operator would: by flying. A lightweight quadcopter built around a Holybro Pixhawk 4 Mini flight controller, four 2300 kV motors, and a 4-in-1 electronic speed controller flew take-off, steady flight, and landing profiles while the controller logged output voltage and current at 10 Hz. Flights spanned take-off weights of 150 to 250 grams and durations of 5 to 15 minutes to cover a realistic spread of missions. After each flight the pack was recharged on a balance charger using the same protocol every time.

Both are lithium polymer packs typical of lightweight UAVs. The larger 2,200 mAh packs produced more flights each because they take more cycles to reach end-of-life.

Because a full discharge damages the pack, remaining capacity was measured on a battery tester only after every tenth flight, with the ten flights in between flown at the same weight and duration. The capacity for each intermediate flight was interpolated linearly between the two surrounding measurements. The paper states that assumption openly, and it is how all 631 flights ended up with a usable label.

Turning flight logs into images

The central idea of the paper is a workaround for data scarcity: instead of training a large time-series model from scratch, reshape each flight's sensor data into a picture and let a vision model that has already learned from 14 million photographs do the feature extraction. Each flight goes through the following steps:

1. Each signal (voltage, current, throttle) is a one-dimensional series of measurements over the flight. It is normalized to a 0 to 255 range, the scale of a grayscale pixel.
2. The series is reshaped into a two-dimensional matrix 112 columns wide and rounded to whole numbers, giving one grayscale image per signal. Resolution is reduced, but the images still resolve voltage changes of 0.05 V and current changes of 0.0075 A.
3. The three grayscale images are stacked into a single three-channel image, one channel per signal, like the red, green, and blue of a photograph.
4. Because flights differ in length, images are padded with black to a common size and resized to 80 by 80 pixels, the input size used for ResNet-50.
5. A ResNet-50 pretrained on ImageNet, with its weights frozen, turns each image into a vector of 2,048 features. No battery-specific training happens at this stage.
6. Small trainable fully connected layers map those 2,048 features to a single number: the predicted remaining capacity after that flight.
7. The previous flight's remaining capacity is fed in as an extra input (autoregression), so the model carries history forward instead of judging each flight in isolation.

Because ResNet-50 does the heavy lifting with weights nobody has to train, the number of trainable parameters is small enough to learn from a few hundred flights without overfitting. Hyperparameters were tuned by sampling 500 candidate configurations with Latin hypercube sampling and keeping the one with the lowest validation error over 2,000 epochs.

Transfer learning between two pack types

The second idea addresses variety. The target was the 1,100 mAh pack, which had the smaller dataset: 4 packs and 203 flights for training, and 1 held-out pack with 86 flights for testing. The 342 flights from the 2,200 mAh packs were used to pretrain the fully connected layers first, then the model was fine-tuned on the 1,100 mAh training data. In effect, the model learned what degradation looks like on one chemistry and size, then adapted that knowledge to another.

The effect shows up in the training curve. Right after the switch to the new battery data the loss ticked up briefly, then fell below the model trained without transfer learning. At the final epoch, training loss with transfer learning was 28 percent lower than without it. For a fleet, that is the property that matters: a model trained on the packs you already own can be adapted to a new pack type with far fewer flights than starting over.

The results

The proposed approach was compared against two models commonly used for battery SoH prediction, a convolutional neural network trained from scratch and a transformer, each tuned with the same hyperparameter search. All three were evaluated on the same unseen test battery:

MSE is mean squared error; MAPE is mean absolute percentage error, the average size of the prediction miss relative to the true remaining capacity. Test set: one 1,100 mAh battery and 86 flights the model never saw during training.

Two things stand out. The transformer, the most sophisticated model in the comparison, did worst on the test set and overfit the training data heavily, which is exactly the small-data failure the paper set out to avoid. And the pretrained vision model alone already beat both baselines by a wide margin; transfer learning then cut its remaining error by a further third.

Since publication, training has continued on a larger dataset, and the test error on the current model stands at 1.57 percent.

What this means when you evaluate a vendor

• It works on data you already have. Voltage, current, and throttle are logged by any modern flight controller. There is no hardware to add and no weight penalty, so any solution that asks you to instrument your aircraft first should explain why.
• It works with small datasets. A few hundred flights on a handful of packs was enough for 2.26 percent accuracy. A vendor that needs years of history before producing a forecast is solving the problem the hard way.
• New pack types are cheap to add. Transfer learning means a model trained on your existing batteries adapts to a new type with a fraction of the flights. Ask how a vendor onboards a pack they have never seen.
• Accuracy should be reported on unseen batteries. The 2.26 percent figure is on a pack the model was never trained on, which is the only number that predicts real-world performance. A headline accuracy measured on the training packs is not comparable.
• Retirement becomes a data decision. Instead of pulling packs at a conservative cycle count, you retire them when the predicted capacity approaches 80 percent, and not before. That is where the spares savings come from.
• It is one model, not the whole product. In Vermilion, a data-driven estimate like this one is delivered with the reasoning trace, calibrated confidence, and physical constraints described on the Technology page, so a maintenance lead sees the evidence behind the number, not just the number.

Limits and what comes next

The paper is careful about its scope, and anyone reading it should be too:

• Only discharge cycles were modelled. Charging data is rarely available from UAV batteries in practice, so it was not collected.
• One airframe, two pack types. Broader validation across drones and chemistries is the obvious next step, and it is what the continued training since publication has been doing.
• Labels between full-discharge checks are interpolated, which assumes degradation is roughly even across ten similar flights.
• Temperature, one of the strongest health signals, was left out on purpose to respect the weight constraint. Where a platform does log it, it should help.

The motor is the next component. The follow-on work carried into Vermilion predicts motor state of health from the accelerometers, gyroscopes, barometer, and GPS a drone already carries, plus camera footage to catch the visual instability a degrading motor produces. The fleet-level picture, including what to demand from any solution before it touches your aircraft, is in the companion guide on predictive maintenance for UAV fleets.

Who did the research, and where Vermilion fits

The paper was authored by Jiarui Xie, Lingchen Kong, and Yaoyao Fiona Zhao of McGill University's Department of Mechanical Engineering; Mohamed Rami Latreche and Elaine Mosconi of Université de Sherbrooke's Department of Information Systems and Quantitative Management Methods; and Sean Smith of Vozwin, the founder of Vermilion. It is part of the PHUMS project established by Vozwin, with UAV design and experiments guided by Fabio Bandera of Vozwin, and it was supported by the MITACS Accelerate program and the Réseau SDG Innovation Network's collaborative R&D projects in digital, intelligent, and sustainable transformation.

Vozwin owns the technology presented in the research, which was carried out through its Labs division. Vermilion is the production platform built on it: incubated in Vozwin's venture studio, now operating as its own company, and available for deployment on edge, cloud, or air-gapped infrastructure. If you want to see what the model predicts from your own flight logs, the Contact page books thirty minutes with a Vermilion engineer.