One production context across MES, ERP and the machines themselves.
Most manufacturers do not have a shortage of production data. They have a historian per plant, an MES that knows work orders but not sensor behaviour, an ERP that knows cost but not cycle time, and a quality system that finds drift after the parts are built. Analytics projects then stall in the gap between them — accurate on one line, unusable on the next.
Why plant AI stalls after the pilot line
A predictive maintenance pilot on one asset class usually works. Rolling it out is where it breaks, and the reason is structural rather than statistical: the second plant names its assets differently, samples at a different rate, ran a different retrofit, and records downtime reasons in a free-text field with local conventions.
The model was never the fragile part. The mapping from raw signals to a shared idea of what a machine, a line, and a work order are is the fragile part, and rebuilding it per plant is why so many industrial AI programmes plateau at proof-of-concept.
The production ontology: modelled once, reused per plant
NeuroCluster resolves MES, ERP, PLC and SCADA tags, sensor streams, quality results, bills of material, and maintenance history into one production context: Factory, Production Line, Machine, Component, Sensor, Product, Work Order, and Supplier as governed objects with stable identity and version history.
Signal-level normalisation happens under that model — unit harmonisation, time alignment across differently-sampled sources, and asset-hierarchy resolution so a spindle is the same component whether the historian, the maintenance system, or the bill of material is describing it. New plants map onto the existing model instead of starting a new one.
- Asset hierarchy resolved across historian tags, maintenance records, and the bill of material.
- Work orders, downtime events, and quality results attached to the specific component they concern.
- Supplier and component genealogy retained, so a quality signal can be traced to a batch and a source.
- Time-aligned sensor context, so a model compares like with like across lines that sample differently.
Intelligence that changes a shift, not a slide
On that foundation the analytics are the well-understood part: bottleneck detection across the line, remaining-useful-life models per component class, quality drift detection against process parameters, energy optimisation per unit produced, and supply risk scoring on the components that would actually stop production.
The value is in the join. Quality drift on line 3 correlated with a specific spindle's vibration signature and its maintenance history is an actionable finding; either signal alone is a dashboard.
The maintenance decision, with the plant lead holding authority
A representative recommendation: pull forward the spindle replacement that is driving the quality drift on line 3. It reaches the plant maintenance lead with the evidence attached — the drift signal, the component history, and the production impact of acting now versus at the next planned window.
The lead decides. Once authorised, the intervention is released into the maintenance window as a scheduled work order in the systems the plant already runs, and scrap rate and unplanned downtime are tracked against that intervention. The loop closes with a measured outcome rather than a claimed one, which is also how a model's real-world precision becomes visible instead of theoretical.
Running where the shop floor actually is
Production data is commercially sensitive, latency-bound, and frequently sitting behind an OT boundary that exists for safety reasons. Streaming it to a hosted AI API is often both a security objection and an operational one.
The platform runs on-premises next to the plant, on a dedicated European tenant, or fully air-gapped, with the same governed agents and evidence in each case. Where connectivity is intermittent, reasoning stays local and only outcomes synchronise.
Where the EU Data Act and Machinery Regulation land
Two regulations changed the industrial data picture. The EU Data Act (Regulation (EU) 2023/2854), applicable since 12 September 2025, gives users of connected products a right of access to the data those products generate — which strengthens an operator's position in getting telemetry out of vendor-controlled equipment, and creates obligations for manufacturers who build connected machines.
The Machinery Regulation (EU) 2023/1230, applicable from 20 January 2027, brings safety components with self-evolving behaviour into scope for conformity assessment. Where AI influences a safety function, documentation and traceability are not optional — which again favours a runtime that records what a system did and why.
Frequently asked questions
Do we need labelled failure data for predictive maintenance?
Less than most teams assume, and rarely as the starting point. Anomaly and drift detection against a modelled baseline delivers value before any failure labels exist, and the approved-intervention loop generates labelled outcomes as it runs. The harder prerequisite is asset and signal context — knowing which sensor belongs to which component on which line — which is ontology work rather than modelling work.
Does this replace our MES or ERP?
No. It reads from them and writes decisions back into them. The MES remains the execution system of record and the ERP the commercial one; NeuroCluster is the layer that gives them a shared model, reasons across it, and routes recommendations through human authority before they become work orders.
Can it run air-gapped on the shop floor?
Yes. Models, retrieval, and the agent runtime deploy inside your boundary, including fully disconnected sites. This is a standard deployment profile rather than an exception, because OT segmentation and plant latency requirements make it the common case in industrial settings.
Who is accountable when a recommendation is wrong?
The named human who authorised the action, which is why authority is explicit in the design rather than implied. Every recommendation carries its evidence, every approval records who accepted it, and outcomes are measured against the intervention — so a systematically wrong model becomes visible in the record and can be withdrawn, rather than quietly eroding trust on the floor.
How do we start?
One line, one asset class, one measurable loss — typically unplanned downtime or scrap. We model that slice, run the intelligence on your historical data to establish whether the signal is real, and only then wire the approval and work-order path.
Keep evaluating
Enterprise process management
Governed process intelligence and the approval loop applied to operations.
On-premise and air-gapped AI
Running the full stack next to the plant, including disconnected sites.
Platform overview
The eight layers — from compute and data to ontology, agents, and evidence.
Multi-agent AI systems
How specialised agents coordinate on operational work without losing control.
Critical infrastructure and resilience
For plants in scope for NIS2: dependency modelling and reporting evidence.
All use cases
Production AI applications across industrial and critical sectors.