Vozwin AI helps aerospace manufacturers, maintenance providers and UAV fleet operators across Canada put machine learning to work on the data their aircraft and operations already record, in two forms: predictive maintenance, which forecasts when a component will need service from its flight logs, and predictive analytics, which forecasts maintenance demand, spares and fleet availability from the same history. The work is grounded in Vozwin's own peer-reviewed research, which predicts drone battery state of health within 2.26 percent from the voltage, current and throttle a stock flight controller logs, with no added sensors. We deliver pilots measured against a baseline, integrate the result with the maintenance and fleet systems you run, and structure the work around Canadian R&D funding.
This page covers where predictive maintenance and predictive analytics pay off in aerospace and UAV operations, what your data has to look like before a project is worth starting, how an engagement runs, and where Vozwin AI, Vermilion and Vozwin Aerospace each fit. If you build or operate aircraft that log their own operation, the first question is what that data can already predict.
Where predictive maintenance and predictive analytics pay off
Maintenance is the natural starting point because the data already exists and the cost of getting it wrong is visible. For UAVs, maintenance can represent more than 50 percent of direct operating costs, and published reliability analyses put the overall failure rate of UAV systems around 25 percent. Cross-industry studies put predictive maintenance 8 to 12 percent below preventive maintenance on cost and up to 40 percent below run-to-failure. The use cases we see most often:
- Predictive maintenance from flight logs. Forecast when a component will need service from the voltage, current, throttle, inertial and GPS signals the aircraft records, so parts are replaced on evidence rather than on a calendar or after a failure.
- Battery state of health. Retire packs when predicted capacity approaches 80 percent instead of at a conservative cycle count, rotate the healthiest packs onto the longest missions, and plan endurance against the real capacity of the pack on the aircraft.
- Motor and airframe health. Predict wear from the accelerometers, gyroscopes, barometer and GPS a drone already carries, with camera footage as a second opinion, instead of adding vibration sensors that can weigh a tenth of a small aircraft.
- Predictive analytics for planning and spares. Forecast maintenance demand across the fleet, plan maintenance windows around it, and order spares against forecast dates instead of a safety stock sized for the worst month.
- Fleet availability. Forecast which aircraft will be serviceable on a given date from predicted component health and scheduled work, so missions and crews are planned against the fleet you will actually have.
The value lands in decisions that change, not in the model. A use case that cannot name the decision it would change, and what that decision costs today, is not ready for a pilot.
What your data has to look like
Most aerospace and UAV operations already have more usable data than they think. The readiness check we run on your logs and records asks five questions:
- Does each aircraft, pack or component have a stable identifier that appears in both the flight logs and the maintenance records? Without it, the signals cannot be joined to the outcomes.
- Are flight logs retained with timestamps, at a consistent sampling rate, for long enough to cover the life of the component you care about?
- Do maintenance records capture what was replaced, when, and why, in a form that can be read as a label rather than free text only?
- Is there enough history? The battery research below reached a 2.26 percent error from 631 flights across two pack types. A few hundred well-labelled flights per equipment type can be enough when the method fits the data.
- Who acts on the output? A forecast that reaches a dashboard nobody checks before a flight changes nothing.
If the answers are mostly yes, a pilot can usually start without new instrumentation. If they are mostly no, the first project is the data pipeline, and we say so before anyone budgets for a model.
How an engagement runs
- Readiness check. Two to six weeks on your flight logs, maintenance records and systems, ending in a ranked list of use cases, the data gaps that would block each one, and a recommended first pilot.
- One component, one decision. The pilot picks a single component or equipment class and the single decision the forecast would change, with the baseline written down: how that decision is made today, how often it goes wrong, and what each outcome costs.
- Build, buy or partner. We review the vendor landscape for the use case, test shortlisted products on your data, and build only when the product does not exist or the data advantage is yours to keep.
- Shadow mode. The model runs alongside the current maintenance program on data it never saw during development, and we compare what it would have done with what actually happened.
- Integration. The output lands in the maintenance, fleet or ERP system your crew already uses, with monitoring, retraining and rollback procedures written down.
- Adoption and handover. Training with the people who sign for the equipment, runbooks for operations, and documentation that explains the decisions, so your team runs it through a bad week without calling us.
Built on peer-reviewed research
Under the PHUMS project, Vozwin worked with McGill University and Université de Sherbrooke, supported by the Mitacs Accelerate program, to predict UAV battery state of health from flight data. The team flew 631 experiments on lithium polymer packs of 2,200 mAh and 1,100 mAh, using only the voltage, current and throttle from each discharge cycle. Each flight's time series was converted into an image so a pretrained ResNet-50 vision model could extract features from it, and what the model learned on one pack type was transferred to the other.
The published result is a test error of 2.26 percent on a battery the model had never seen. Continued training on a larger dataset since publication has brought that to 1.57 percent. Motors are the next component, predicted from the sensors a drone already carries. Both guides linked below walk through the method and what it means for a fleet.
Where Vozwin AI, Vermilion and Vozwin Aerospace fit
Vozwin funded and owns the battery and motor state-of-health research described above, carried out through its Labs division. Vermilion is the production platform built on it: a physics-informed reasoning model, incubated in Vozwin's venture studio and now operating as its own company, production-ready and available for deployment on edge, cloud or air-gapped infrastructure. Vozwin Aerospace works the aircraft side, building UAV platforms that log these signals from the first flight.
Vozwin AI is the part of the group that helps an aerospace or UAV operation get from interest to a working pilot and from pilot to production: the readiness check, the build, buy or partner call, the data pipeline and the integration with your maintenance systems, the pilot design and baseline, and the adoption work with your crew. We are vendor-neutral in that work. When Vermilion is the right platform we say so, and when a configured product or a custom build fits better we say that instead.
Constraints we design around
- Weight and power. Nothing we propose adds hardware to the aircraft unless the data already on board has been shown not to carry the signal.
- Approved maintenance programs. Predictions inform the decisions your maintenance program allows and document the evidence behind them. They do not replace the program approved for the aircraft.
- Security and deployment. Defence, public safety and regulated operators often need on-premises or air-gapped deployment, and the architecture is chosen to fit that from the start.
- Bilingual operations. Engagements run in English or French, including training and documentation for crews in Quebec and across Canada.
Canadian funding that applies
Predictive maintenance and state-of-health projects involve genuine technical uncertainty, which is what Canada's main innovation programs are built to support. SR&ED tax credits apply to eligible experimental development, such as building a model for a component where no proven approach exists. NRC IRAP funds part of the salary and contractor costs on eligible innovation projects for small and mid-sized businesses. Mitacs Accelerate co-funds research internships with Canadian universities, which is how the battery research above got access to university-grade modelling expertise. Most provinces add their own technology adoption and R&D support that stacks with the federal programs.
Structure the project around the programs before work starts, not after. Eligibility depends on how the work is scoped and documented, and we scope engagements with that in mind.