What is Industrial AI?
Industrial AI is the application of artificial intelligence, mainly Machine Learning, to data from sensors, machines and control systems in factories, process plants and energy and water utilities, with the aim of predicting failures, monitoring quality and optimising industrial processes. Unlike AI in office applications, Industrial AI works with time series, vibration signals and camera images from physical installations, and every error can cause production losses or danger. That is why, in practice, Industrial AI mainly advises, while the PLC and the safety system keep the actual control.
In an OT context, Industrial AI helps optimise production, predict failures and reduce energy consumption, without losing sight of the Safety and reliability of the process.
🧠 What does Industrial AI do in a plant?
Industrial AI turns raw process data into a prediction, classification or recommendation. The six most common applications are:
- Predictive Maintenance - predicting when a pump, motor or bearing needs maintenance, based on vibration, temperature and current draw
- Anomaly Detection - recognising abnormal behaviour, such as a leak, a sticking valve or suspicious network traffic
- Process optimisation - advice on setpoints for maximum yield or minimum energy and raw material use
- Quality control - visual inspection with Vision AI and cameras (Vision), or a soft sensor that estimates a product quality you would otherwise only measure later in the lab
- Asset health monitoring - continuously tracking the condition of installations, a form of condition monitoring
- Digital Twin models - a virtual model of the physical system that is updated with real-time data and lets you run scenarios
Most models are classic machine learning (regression, decision trees, clustering). Deep Learning is used mainly for image and sound analysis. Generative AI and AI agents form a newer layer that mostly supports operators and maintenance teams; see Agentic AI in OT.
🔍 What are examples in OT environments?
| Use case | Data | What the AI does | Typical result |
|---|---|---|---|
| Pump performance analysis | Pressure, flow, current, vibration | Detects cavitation or declining efficiency | Maintenance before failure instead of after |
| Visual quality control | Camera images per product | Spots hairline cracks, stains, wrong labels | Rejects sorted out automatically |
| Energy savings | Temperatures, pressures, load | Optimises cooling or compressed-air installations | Lower consumption at the same output |
| OEE analysis | Machine states, downtime reasons | Finds patterns in downtime and micro-stops | Targeted improvement actions for OEE |
| Soft sensor | Process variables from the Historian | Estimates a lab value, such as viscosity | Faster correction than waiting for the lab result |
| Cyber anomaly detection | Network traffic in the OT zone | Recognises abnormal behaviour of a PLC or SCADA system | Early warning of misuse |
🔄 How does Industrial AI differ from classic control engineering?
Classic control engineering, such as a PID controller or MPC within APC, works with an explicit model that engineers derive from physics and step tests. Industrial AI learns a model from historical data. The two complement each other: AI supplies an estimate or prediction, for example, while the control loop executes it in a stable, deterministic way.
| Characteristic | Classic control (PID, MPC) | Industrial AI (ML model) |
|---|---|---|
| Basis | Physics and process identification | Patterns in historical data |
| Behaviour | Deterministic and analysable | Statistical, dependent on training data |
| Place in the architecture | PID in the PLC or DCS (level 1), MPC usually on an APC server (level 2) | Usually an edge server at level 2–3 or higher, or in the cloud |
| Process changes | Retuning by an engineer | Retraining with new data |
| Explainability | High: every control action can be traced | Limited; needs additional explanation techniques |
| Role in safety functions | Not the control loop itself: safety functions run as deterministic logic in an SIS under IEC 61511 and functional safety | None; advisory at most, outside the SIS |
An intermediate form is the hybrid model: a physical model or digital twin that an ML component adjusts on the basis of current measurements. A digital twin therefore does not learn by itself; it only does so once a data-driven model is attached to it.
🔧 How does Industrial AI connect to existing systems?
| System | Role in the AI architecture |
|---|---|
| SCADA | Source of current process data and alarms |
| Historian | Time series for training and predictive analysis |
| MES | Link to orders, batches and quality data |
| Edge Computing | AI inference close to the sensor or machine, without depending on the internet |
| OPC UA and Unified Namespace | Standardised access to data with context |
| Digital Twin | Simulation model of the physical process |
In practice, data travels from field level via the historian or an MQTT broker to an edge server or the cloud. Training usually happens centrally, while inference increasingly runs locally at the edge. Recommendations go back to the process through the operator, or through a bounded write action that is secured as a separate conduit in line with the Purdue Model and IEC 62443.
Collaboration matters as much as technology. An Industrial AI project needs data science, process knowledge and OT operations at the same time; without process engineers and operators on the team, you end up with models that are statistically sound but unusable in practice. That makes Industrial AI a typical example of OT convergence.
🧭 How do you approach an Industrial AI project step by step?
- Pick one measurable problem - for example unplanned downtime of one pump type or the reject rate of one line, with a baseline measurement and a fixed team of data scientist, process engineer and operator.
- Check the data - are the signals available, at sufficient frequency, with timestamps and context? Vibration analysis needs measurements in the kHz range, whereas a historian often stores process variables per second or per minute.
- Label the history - link failures, maintenance notifications and quality results to the time series. Without labels, only anomaly detection remains.
- Build and validate offline - test the model on data it has not seen, and let process experts judge the results.
- Run in shadow mode - the model predicts alongside operations, but nobody acts on it yet. This shows you how many false alarms it produces.
- Integrate as advice - show the recommendation on the SCADA screen or as a work order; changes to the process go through change management.
- Monitor and retrain - process changes, new raw materials or a recalibrated sensor cause model drift; plan periodic retraining and clear ownership.
For condition monitoring, ISO 17359 (setting up a monitoring programme), ISO 13374 (processing model from measurement to diagnosis and prognosis) and ISO 13381-1 (prognosis of remaining useful life) provide a useful framework.
| Data type | Typical source | Common AI technique |
|---|---|---|
| Vibration and sound | Accelerometers, microphones | Spectrum analysis with ML classification |
| Process variables | Historian, SCADA | Regression, anomaly detection on time series |
| Images | Industrial cameras | Deep learning (convolutional networks) |
| Events and text | Alarm logs, maintenance notifications | Clustering, language models |
| Network traffic | SPAN port in the OT zone | Baselining and behaviour analysis |
⚠️ What are the risks of Industrial AI?
- Data quality - missing timestamps, frozen sensor values or wrong units produce an unreliable model.
- Erroneous recommendations - bad advice can cause production losses or damage, so keep a human in the loop.
- Model drift - a model that worked well last year can silently degrade after a process change and produce false or missed alarms.
- Over-automation - operators who trust AI blindly lose the ability to intervene without it.
- Lack of explainability - advice that operators cannot trace back is either ignored or followed blindly; so also show which signals led to the recommendation.
- Attacks on AI - data poisoning (manipulated training data), evasion (input that misleads the model) and theft or tampering of models. MITRE ATLAS describes these techniques, modelled on MITRE ATT&CK.
- New access paths - a cloud connection for AI can open a route into the OT zone; preferably use one-way traffic, for example via a data diode.
On 3 December 2025, CISA and Australia’s ASD’s ACSC, together with partners including the Dutch NCSC, Germany’s BSI and the NSA, published Principles for the Secure Integration of Artificial Intelligence in Operational Technology. The core message: understand the risks, test whether AI in OT is really needed, establish governance, and ensure human oversight and a fallback option. According to the guidance, AI such as language models almost certainly does not belong in safety decisions; the SIS stays deterministic and separate from AI.
🇪🇺 Which rules and frameworks apply in the Netherlands and the EU?
| Framework | What it means for Industrial AI | Status |
|---|---|---|
| AI Act (EU) 2024/1689 | Risk classes; high risk when AI is a safety component | AI literacy since 2 February 2025; high risk for critical infrastructure from 2 December 2027, machinery from 2 August 2028 |
| Machinery Regulation (EU) 2023/1230 | Requirements for machines with self-learning behaviour | Applies from 20 January 2027 |
| Data Act (EU) 2023/2854 | Users’ right to data from connected machines | Applies since 12 September 2025 |
| ISO/IEC 42001:2023 | Certifiable management system for AI | Published December 2023 |
| NIST AI RMF 1.0 | Risk approach with Govern, Map, Measure and Manage | Voluntary, since January 2023 |
Most Industrial AI applications, such as predictive maintenance or energy optimisation, fall into the minimal-risk class under the AI Act. The Data Act often matters more for Industrial AI: a plant can request the data from its own machines and have it analysed by an independent party. In the Netherlands, a European AI factory is also coming to Groningen: EuroHPC selected the proposal from partners including SURF, TNO and AIC4NL in October 2025, with a total investment of over 200 million euros. The supercomputer is expected to be operational in early 2028 and is meant to give start-ups, SMEs, researchers and public bodies access to computing power for training AI models.
❓ Frequently asked questions
What is the difference between AI and Industrial AI?
Industrial AI is AI tailored to industrial environments, using data from sensors, machines and controllers rather than text or customer records. Industrial AI has to account for real-time requirements, process safety and the long service life of installations. Errors therefore have physical consequences, which places stricter demands on validation and human oversight.
Can Industrial AI control a PLC directly?
Industrial AI can technically write setpoints to a PLC, but common practice is that the AI advises and an operator or a bounded control loop decides. A direct connection to the safety system is never acceptable. The international guidance on AI in OT from December 2025 advises against letting AI make safety decisions.
How much data does Industrial AI need?
Industrial AI usually needs months to years of historical process data, including examples of the failures or deviations you want to recognise. For rare failures, those examples are often missing; you then start with anomaly detection, which only has to learn normal behaviour. Data quality and context matter more than sheer volume.
Does the AI Act apply to Industrial AI?
The AI Act applies to Industrial AI, but usually in the minimal-risk class, as with predictive maintenance or process optimisation without a safety function. If AI becomes a safety component of a machine or of critical infrastructure, it counts as high risk, with requirements from 2 December 2027 or 2 August 2028. The AI literacy obligation has applied since 2 February 2025 to every organisation that uses AI.
Why does Industrial AI often run at the edge?
Industrial AI often runs on an edge server because inference is then fast, even with an unreliable internet connection, and process data does not have to leave the site. That reduces the attack surface and the dependency on a cloud provider. Training models, however, usually still happens centrally or in the cloud.
What is model drift in Industrial AI?
Model drift is when an Industrial AI model gradually or suddenly performs worse because the process changes, for example through new raw materials, wear or the recalibration of a sensor. The result is more false or missed alarms. You prevent this by monitoring performance continuously and retraining the model periodically.
📌 In summary
Industrial AI learns from process data to predict failures, monitor quality and optimise processes, but it advises and stays outside safety functions. Success requires good data, a measurable problem, integration via historian, edge and OPC UA, monitoring for model drift and security in line with IEC 62443, within the frameworks of the AI Act, the Data Act and the Machinery Regulation.