For industrial ML teams

From sensor data to a model advantage.

Machines that operate across regimes, faults that develop slowly, false alarms that cost real money. Ensue runs coordinated research on your telemetry against the metrics that matter on the floor: prediction error, early detection, cross-unit transfer, and the hardware budget you actually deploy on.
How it works →

Research areas

The problems we work on.

Remaining useful life

Sensor-based failure prediction, evaluated on official splits under the same protocol as the published literature.

Early detection without false alarms

Separate “is it degrading?” from “when will it fail?”, so sensitivity doesn’t come at the cost of trust.

Cross-unit transfer

Condition normalization and per-unit baselines, so a model trained on some machines holds up on the rest.

Multi-condition robustness

Operating regimes change the data. Models that survive regime shifts instead of memorizing one of them.

Edge deployment constraints

Millisecond inference on plant hardware. Complexity constrained on purpose, to transfer instead of memorize.

Feature discovery from raw telemetry

Multi-scale trend detection and automated feature selection across the sensors you already log.

The engagement

How an engagement runs.

  1. 01

    Bring your telemetry

    Sensor streams, failure history, maintenance logs. It stays in your repo, on your infrastructure when required.

  2. 02

    Define the operating constraints

    The metric, the cost of a false alarm, and the hardware budget. Those constraints shape every experiment.

  3. 03

    Experiments run continuously

    Architectures, normalizations, and feature treatments swept against your evaluation on dedicated compute.

  4. 04

    Improvements arrive as pull requests

    Every keep is re-verified on held-out data and lands in your repo with the experiment trail attached.

Proof

Verified, not vibes.

22–34%better than published deep-learning methods on NASA’s C-MAPSS turbofan benchmark under multi-condition evaluation
<1msinference, trains in under five minutes on a laptop, no GPU required
3 datasetsvalidated on real sensor data from NASA, DOE/LBNL, and the PHM Society. No synthetic data.

Take it with you

The turbofan RUL report, as a PDF.

The NASA C-MAPSS results in full: the six changes that carried them, head-to-head tables against published deep learning, and validation on real sensor data.

You’ll also get the occasional research note from Ensue. No spam, ever.

Tell us the machine and the metric.

Bring the telemetry you already collect and the failure that costs you the most. We’ll scope the evaluation on the first call.