TECHNOLOGY

Learning, moved onto the device.

Most machine-learning systems learn somewhere else and are then shipped as a fixed model. RailMind puts two parts on the device instead: an engine that keeps learning while the device runs, and a Neuromodulation Layer that decides when it may learn.

Any stream of data

Machines
Structures
Cameras
Robots and devices

Neuromodulation Layer

The gate on learning. The system raises the signal; your operators decide.

Learning engine

Learns what normal looks like, on the device

What it has learned

Kept on the device. Changes are shown to your team.

Schematic: the two parts on one device. Inputs are examples.

What happens on the device

Attach

The engine goes where the data starts: on a machine, a structure, a camera or another maker’s device. It starts with nothing learned.

Learn

A picture of what is normal builds up from that source’s own data, on the device, while it runs.

Notice

When the data moves away from what it has learned, the change is shown to your team.

Decide

Before maintenance or a planned change, learning is paused, so the change is not absorbed as the new normal. The Neuromodulation Layer is that gate: the system raises the signal, and your operators decide when learning pauses and resumes.

CONTROLLED PLASTICITY

A device that learns needs to know when to learn.

An engine that keeps learning adapts to what normal means for each source it watches. But not every change it sees is worth learning.

Schematic animation.

The Neuromodulation Layer

The Neuromodulation Layer is the gate on learning: it controls when learning happens. The system raises the signal, and your operators decide when learning pauses and resumes — much like the neuromodulators that regulate learning in the brain.

A model that keeps getting to know this machine, while faults stay faults.
Learning that stops and resumes only on an explicit operator decision, so every change has an owner.
An approach in line with the direction regulators are setting for systems that update themselves.

Neuromorphic — but not all neuromorphic is the same.

Comparison of monitoring approaches
	RailMind	Fixed alarm limits	Cloud models	Typical edge AItrained beforehand
Learns each asset's own normal	Yes	No	Partly	No
Keeps learning after installation	Yes	No	Partly	No
Data stays on site	Yes	Yes	No	Yes
No labelled fault data needed	Yes	Yes	Partly	Partly
Your team decides when learning pauses	Yes	Not applicable	No	Not applicable

“Partly”: cloud models are usually retrained centrally, in batches, not on the device; some edge AI is trained on normal data only.

What sets it apart

Learns on the device

The model forms on the device, from the data of whatever it watches.

No training data up front

No labelled faults and no data-collection phase before it is useful.

Keeps learning, under your control

It follows the asset as it ages and its load changes. Your team pauses and resumes learning.

Built for small hardware

Designed from the start for low-power microcontrollers, not shrunk down from a data-centre model.

Where the engine runs

Today

Embedded gateway

On sites that already have compute. No additional hardware needed.

Today

Microcontroller

On the asset itself, inside a small device.

In silicon

The engine as a part, inside the sensors and devices others build.

How it fits into a site

Four pieces: sensors on the asset, a node that learns it, a hub for the site, and the platform you already use.

System overview: sensors and assets connect to Edge Nodes, cameras connect to the Edge Hub, Edge GTS adds time and position, the hub feeds your platform

Demo video coming soon.

Where the model lives

On the device, next to the asset it describes. Readings can be sent on to your systems if you want them there; the learning itself does not depend on a connection.

Who is in control

Your team. Learning pauses and resumes when you decide.

ENGINE

Building your own devices?

License the engine and put the learning inside your own product.

Talk to us about a licence

See the devices, or start with one machine.

Products Discuss a pilot