Home/Industries/Industrial Equipment & Machinery/Electric Cylinders
Industrial EquipmentForce EstimationValidated on Public Data

Estimating the axial force on a ballscrew without adding a sensor

An edge AI model that estimates the axial force on a ballscrew from the motor current and speed a drive already reports. No force sensor, no cloud connection, running on the existing controller.

Context

In an electric cylinder, whatever the machine pushes against ends up as axial force on the ballscrew and its bearings. In press-fit, tightening and pressing work that force is the process: it decides whether the part is good. It is also where wear and a failing lubricant show themselves first.

Fitting a load cell to a production machine is another matter. Cost and the space it takes rule it out of most designs, so a quantity a test rig measures directly stays invisible on the equipment actually running. Motor current is the usual stand-in. But current tracks torque, and between torque and axial force sit the inertia and the compliance of the mechanism; the two are not related by a constant.

What we tested

We built a model that estimates axial force from the signals a drive already holds: the individual phase currents, the active current, and the speed. Nothing was added to the machine. The model is built on the family of lightweight machine-learning techniques we work with across our projects (reservoir computing).

Validation used a public dataset in which electric cylinders were run continuously to failure with every sensor recorded throughout. We applied the same procedure to three cylinders of the same type. Each model saw only the early-life interval of its own cylinder, with no data pooled across units, and accuracy was measured on a later interval used neither for training nor for tuning.

What it shows

The force waveform could be reproduced on the held-out interval, and at the same level on all three cylinders: the quantity is available without a load cell in the machine.

The model this takes is under a kilobyte and runs on the existing controller as it is. No GPU, no cloud connection.

An internal validation on publicly available data, not a result from a production deployment. Late in service life the estimate stops tracking the measurement; the model is trained on healthy behaviour, so it does not carry over to a machine already close to failure. Anomaly detection and remaining-life estimation were not part of this work.

This may apply to your equipment

If the pushing force decides your part quality or your service intervals and a load cell is impractical to fit, the same idea may be worth exploring: electric cylinders, electric presses, press-fit and tightening stations, transfer and positioning axes. The signals it needs are ones your drive already reports. A short conversation is normally enough to tell whether your case fits.

The inverse approach, pulling out the force a motion command does not account for, is described on the servo axis torque page.

Have a similar challenge?

We're accepting new proof-of-concept projects throughout 2026. If you have a real-time prediction problem that needs to run on existing edge hardware, we'd be glad to talk it through with you.

Discuss a Proof of Concept
Imprint
Entrox Systems GmbH (in formation)
Pelkovenstraße 71, 80992 Munich, Germany
Managing Directors: Sebastian Baur, Daniel Köglmayr
Email: info@entrox-systems.com · Phone: +49 170 6936478
Commercial Register: Amtsgericht München, HRB number pending
VAT ID: pending
Responsible for content pursuant to § 18 (2) MStV: Tamon Nakano, Pelkovenstraße 71, 80992 Munich, Germany