Artificial Intelligence
for Industry (Edge AI)

Industrial artificial intelligence is the use of machine-learning models on the data a plant already produces (PLCs, sensors, cameras and systems) to detect, count, predict and explain what happens in production, with the latency and security a physical process demands. Unlike consumer AI it does not answer general questions; it answers plant questions. How many units left line 3 this shift? Will this motor fail before the weekend? Why did OEE drop on Tuesday?

Three traits set it apart. It works on live process data, not internet text. It runs where the machines are, at the edge or on a plant server, because a decision that takes half a second arrives late. And it has to integrate with what already exists (SCADA, MES, ERP) instead of replacing it. In mid-sized plants running equipment from several generations, that third point usually decides whether a project ships or stays a pilot.

A plant's problem is not a lack of data. Its data lives in five pillars that do not talk to each other:

• Machines. PLCs, sensors and cameras emit signals all shift long, in protocols management cannot read: Modbus, OPC-UA, Profinet, LoRa.
• Systems. SCADA, MES and ERP each know their part and none knows the whole. Between them, someone types it by hand.
• People. What the machine never records: the shift, the spreadsheet, the WhatsApp thread. No evidence, and it is gone at shift change.
• Management. Reports and decisions that arrive at month end, not with today's shift.
• Knowledge. The owner's or plant manager's judgement, which is never written down.

What is missing between the five is context: what happened, where, when and with what evidence. A model that sees only one pillar produces alerts nobody can explain, and an alert nobody explains gets muted. That is why useful industrial AI starts by connecting the pillars and only then by training models. It is the idea behind the industrial second brain: a plant memory that people and agents query with evidence.

• Predictive maintenance. Vibration, current, temperature and sound from motors, gearboxes and pumps. The model learns each asset's normal signature and warns days ahead instead of minutes. See predictive maintenance and the IoT predictive maintenance guide.
• Machine vision for quality and counting. Ordinary IP cameras, often the security cameras already on the wall, processed on site with networks such as YOLOv8: YOLOv8 for defect detection and counting and automated counting.
• OEE and production. Counting per sensor and per recipe, stops per machine, downtime explained. See the OEE guide.
• Energy. Consumption per batch or per unit produced, and electrical anomalies that anticipate mechanical failure.
• People safety. Local vision for exclusion zones, lifelines and PPE, with no footage leaving the plant.
• Knowledge and agents. Manuals, logbooks and meeting notes queried in plain language with evidence (RAG), and agents that explain a deviation instead of just flagging it. See agentic AI in manufacturing and physical AI.

Four customers on a monthly subscription, in production. None had to replace equipment: the platform runs on cameras, sensors and machines already installed, or on simple counting sensors.

• Grupo PROA, poultry. Egg counting by vision on 19 existing security cameras across four farms, one industrial PC with a GPU per farm, integrated in real time with SAP S/4HANA. It replaced manual counting, with accuracy measured against that count at 99 %, and has been running for two years.
• RGM Mallas de Alambre, metalworking. Six counting sensors on the mesh machines with per-product recipes. Output per machine and per recipe available to supervisors with no manual entry, integrated with the ERP. In production since April 2024.
• Intercos, cosmetics. MES on packaging lines with counting sensors: units per shift, stops per machine and Excel export.
• Ditzler, food. Calibrated temperature monitoring in food production.

Details in real cases.

Not all industrial AI has to run at the edge. The practical rule is set by latency and by data:

• Edge. When the decision feeds the process or a high-speed camera: inference under 5 ms and operation without internet.
• On-premise. A server in the plant to consolidate lines, run the production module (OEE) and the second brain. Data never leaves the site.
• Cloud. Corporate reporting, plant-to-plant comparison and models retrained on months of history.

The customer decides, not the vendor: the same platform runs in all three. The most common policy is process data inside the plant and only aggregated indicators outside. That keeps IT simple: no inbound ports to the internet, IEC 62443 as the framework and camera footage handled under Chile's data-protection law. More in edge computing for industrial networks.

1. Weeks 1 and 2. Pick one line and one question: a count, a recurring stop, a failure that keeps coming back. Inventory what already exists: cameras, PLCs, sensors, systems.
2. Weeks 3 to 6. Connect over Modbus, OPC-UA or MQTT, install the edge device and capture data. The model is tuned on that line's data, not on generic data.
3. Weeks 7 to 10. Run in parallel with the manual count or log and measure the difference every day.
4. Weeks 11 and 12. Close with numbers: accuracy against the manual record, hours freed, stops explained. Decide whether to extend to other lines.

The success criterion is written before the pilot starts. If there is no number measured against the real process at the end, there was no pilot.

• Starting with the model instead of the data. Without a reliable count there is no prediction worth having.
• Replacing equipment that works instead of connecting it.
• Sending everything to the cloud and discovering the latency and the bandwidth bill afterwards.
• Alerts without context. A model that says “anomaly” without where, when or evidence is muted within two weeks.
• Measuring the pilot with the vendor's metric instead of the process's.