Industrial Edge AI: When to Run AI Next to the Machine

An AI model that passes its cloud trial is not automatically ready for the production line. Once live, new questions appear: does the answer arrive before the product has passed, what happens when the site loses internet, and what does it cost to stream camera video to a server all day?

Industrial edge AI handles some of those cases by running the model on a device next to the machine rather than on a distant server. It is not always the best choice. This article covers when AI should run on site, when the cloud is enough, and the upkeep costs that are easy to miss.

What industrial edge AI is

Using machine learning has two stages. Training is when a model learns from historical data; it needs plenty of data and computing power, but happens only now and then. Inference is when the finished model judges new data, such as scoring risk every minute or checking every product image.

Edge AI means inference runs close to the data source. Where inference can sit, from nearest to farthest:

• On the device or machine itself, such as a smart camera or a compute module in the machine panel.
• On an industrial computer at the line, serving several machines or cameras.
• On an on-premise server, on the plant network but away from the machines.
• In the cloud, on servers in a data centre reached over the internet.

Training almost always stays central, where historical data from many machines comes together; usually only inference moves to the edge.

Four questions that decide it

1. How fast must the decision be made?

If the answer must trigger something within one machine cycle, such as rejecting a defective part, the round trip to the cloud becomes a problem. As an illustration, a conveyor moving at 1 metre per second with 0.5 metres between camera and reject gate leaves half a second to capture the image, run the model, and fire the actuator, and internet latency is never guaranteed. Decisions made over hours or days, such as maintenance scheduling, do not need the edge.

2. Can the site's connection be relied on?

Plants in industrial estates usually have stable links; pump stations, plantations, mines, and remote substations often rely on a patchy cellular signal. Ask what happens if the model is unreachable during an outage. If only a report is delayed, the cloud is fine. If on-site decisions must continue, inference has to live on site. Connectivity options are compared in our guide to choosing IoT connectivity.

3. How much data is there?

A temperature sensor sending one value a minute produces 1,440 values a day, cheap to send to the cloud. A raw vibration signal sampled 10,000 times a second (an illustration) produces 864 million values a day per axis, and camera video is far larger still. For such data, the edge processes the raw stream on site and sends only results: scores, events, or summaries such as RMS vibration.

4. Who will look after the hardware?

A cloud model lives in one place: updated once, monitored from one dashboard. Edge models are spread across many devices that each need updating, securing, and replacing, often in hot, dusty rooms. These costs rarely appear in proposals, yet decide whether the system still runs two years on.

Data policy matters too. Some companies do not let video of work areas or certain process data leave the site, which makes on-premise inference a requirement.

Examples: where the model usually runs

A general guide, not a rule:

Use case	Where inference usually fits	Main reason
Rejecting defects on a fast conveyor	Edge, at the line	Latency and image volume
Anomaly detection on high-frequency vibration	Hybrid: summaries computed at the edge, trend model in the cloud	Raw data volume
Predicting maintenance needs from daily trends	Cloud	Decisions measured in days
Energy or production forecasting	Cloud	Needs history and data across sites
CCTV analytics for area safety	Edge or on-premise	Video volume and privacy policy
Comparing performance across plants	Cloud	Data must come together in one place

The common pattern: train centrally, run at the edge

The choice is rarely all or nothing. A common pattern:

1. Historical data is gathered on a central platform, and the model is trained there.
2. The tested model is packaged and sent to edge devices, often in a lighter version that fits smaller hardware.
3. Edge devices run inference and send results (scores, events, some sample data) back to the central platform.
4. Model performance is monitored centrally. If accuracy drops as materials, products, or machines change, the model is retrained and redeployed.

Fast decisions stay on site; knowledge about every machine stays in one place. A model often deployed this way is covered in machine anomaly detection with machine learning.

Costs and upkeep that get overlooked

• Hardware built for the plant floor. Office computers often cannot cope with plant heat, dust, and vibration.
• Remote updates. New models and software must reach every device, with a way to roll back a bad release.
• Device health monitoring. An edge device that fails silently stops making decisions without anyone noticing.
• Security. Every extra device on the OT network is another way in. See OT and industrial IoT cybersecurity.
• Clear ownership. Decide up front who deals with a broken edge device: IT, maintenance, or the supplier.

Things to keep in mind

• The edge does not fix bad data. A weak model in the cloud is still weak at the edge.
• Edge models are often slimmed down, so retest their accuracy rather than assuming it.
• Safety functions stay in protection systems and PLCs. AI should advise or trigger non-critical actions, not replace safety interlocks.
• Prove the value first. Unless latency is the core requirement, test the model in the cloud and buy edge hardware once it has proved useful.

Frequently asked questions

Is edge AI always faster?

Not always. The edge removes network travel, but edge hardware is usually weaker than a server, so a large model can take longer overall. What the edge offers most reliably is predictable timing.

Does edge AI need a GPU?

It depends on the model. Models for sensor data such as temperature and vibration often run fine on an ordinary processor. Image and video models generally need an AI accelerator to keep up.

Can we start in the cloud and move to the edge later?

Yes, and it is often safer. A model proven in the cloud can be packaged for the edge if latency needs and model size are noted from the start.

If inference runs at the edge, does data still go to the cloud?

Usually yes, but much less of it: inference results, summaries, and sample data for retraining. Without it, model performance cannot be monitored.

Industrial edge AI with INCLUDE

Through our predictive analytics and machine learning services, the INCLUDE team starts from the decision you want to improve and a data review, then trains and tests the model on your historical records. Models can run in the cloud, or on an edge device when decisions must be fast or connectivity is limited.

Wherever it runs, model output appears on the INCLUDE platform as scores, forecasts, and alarms alongside your sensor data, accuracy is tracked, and the model is retrained when operating conditions change. Where data must stay on site, the Enterprise plan offers on-premise or hybrid deployment and custom AI models; see the pricing page.