Sector · Manufacturing

AI in manufacturing. Predictive maintenance on your shop floor.

We build AI solutions that turn sensor data from your machines into concrete maintenance actions, before a breakdown occurs. With integrations into your existing CMMS, PLC and ERP. Not a lab experiment, but production-ready on the shop floor.

Use casePredictive maintenance

Use caseQuality control

Use caseYield optimisation

Use caseEnergy & safety

Book a shop-floor audit See our work

The Dutch manufacturing industry in figures.

~71.000

Industrial manufacturing companies in the Netherlands

~830.000

People working in industry

~12%

Share of industry in Dutch GDP

61%

Cite unplanned downtime as the biggest OEE risk

Source: CBS Statline Industry 2024, FME Manufacturing Monitor.

Standard predictive maintenance platforms rarely fit your machine park.

The major PdM platforms — Siemens Industrial Edge, GE Predix, Schneider EcoStruxure, Rockwell FactoryTalk Analytics, PTC ThingWorx, SAP Predictive Maintenance — work well in a homogeneous single-vendor park. The reality on most Dutch shop floors is different: a Siemens line next to an Allen-Bradley cell, a Mitsubishi press in the corner, and a few older machines without native connectivity.

Result: you buy licences that cover only part of your park, get dashboards the engineer does not read, and the valuable correlations between machines and process parameters disappear into vendor silos. The ROI on a standard package often disappoints because the model a vendor supplies was never trained on your products, your cycles, your operators.

A custom AI solution approaches this differently: one data layer on top of all your machines, models trained on your own history, and an interface that the maintenance engineer on the floor — not the IT manager behind a desk — actually uses. See also our approach to custom manufacturing software and enterprise AI implementation.

In practice the difference lies in the details: an engineer who sees with each alert which component is likely to fail, which parts are in stock, and what work-order history already exists for that exact asset. That is where adoption is won or lost. A dashboard opened only by the site manager is a dead dashboard. What you need is a tool that sits in the hands of both the operator and the engineer, on tablet or phone, with the right information at the right moment.

For OEMs and machine builders, there is also a commercial layer: predictive maintenance sold as a service alongside every new machine, with multi-tenant data isolation per end customer and SLA reporting. This opens a recurring service revenue stream alongside one-off machine sales. We build that service layer, including the customer portal, user management and billing integration.

AI use cases that prove themselves on the shop floor.

Predictive maintenance is usually the first step. After that, quality control, yield optimisation and energy and safety applications open up on the same data infrastructure.

Predictive maintenance

The top use case for AI in manufacturing. Using sensors for vibration, temperature, motor current and pressure, you build a continuous picture of the health of every asset. A well-trained model detects deviations that an operator would only notice when it's already too late, and links them to a Remaining Useful Life prediction that your work scheduling can act on.

We design the complete stack: sensor selection, edge inference in the control cabinet, anomaly detection models at cloud level, integration with your CMMS maintenance system and a mobile app for the engineer. We train the model on your history, validate it alongside existing inspection rounds, and only go live once the false-positive rate is acceptable.

Alongside predictive maintenance, we build use cases for quality control (computer vision on the assembly line for defect detection), yield optimisation (tuning process parameters based on AI recommendations, with the operator in the loop), demand forecasting (which product, which batch, which volume), energy optimisation (HVAC plus machine consumption), worker safety (camera vision for PPE checks and hazardous actions), supply chain AI, robotics (cobots and autonomous mobile robots), digital twin with AI simulation and document AI (CE technical files, quality reports, ISO work instructions). In many cases, the data infrastructure you set up for predictive maintenance is the same layer these follow-on cases run on, so that foundation pays for itself twice.

Vibration and motor current monitoringDetection of bearing, balance and alignment issues before audible damage occurs.
Remaining Useful Life modelsAn RUL estimate per asset that planning can act on.
CMMS integration with automatic work ordersA prediction triggers a work order and parts pick directly.
Mobile app for engineersWork order, machine history and root cause on one screen on the shop floor.

Built on the standards that matter in industry.

Compliance is not a separate audit step. From the first sprint, we work within the framework of Industry 4.0, ISO standards, the EU Machinery Regulation and the AI Act.

Industry 4.0 / RAMI 4.0

Asset Administration Shell

We model machines according to the Reference Architecture Model Industrie 4.0, with the Asset Administration Shell as the interoperability layer. This keeps your data foundation vendor-independent and extensible.

ISO 22301 — BCM

Business Continuity Management

Predictive maintenance is a continuity tool. We document the impact of model failure, provide fallback procedures, and link critical alerts to your BCM plan.

ISO 27001

Information security

OT data leaves your factory only through encrypted tunnels. Access control, audit logs and key management are set up for ISO 27001 certification, optionally with an external audit partner.

EU Machinery Regulation 2027

Safety of AI components

The new EU Machinery Regulation treats AI software that affects machine safety as a safety component. We map out risk classification, conformity assessment and the CE technical file from the scoping phase.

EU AI Act

High-risk classification for safety AI

Safety-critical AI (worker-safety camera detection, machine safety decisions) falls under high-risk. Data quality, transparency, human oversight and log retention are part of our defaults.

Seamlessly connected to your existing OT and IT stack.

We connect to the controls, sensors and software your factory already uses. No vendor lock-in: our layer complements what you have and stays manageable by your own team.

OPC UA

Machine-to-cloud protocol

MQTT

IoT messaging

Modbus

Legacy PLC integration

Azure IoT

Cloud aggregation

AWS IoT

Cloud aggregation

Siemens S7

PLC family

Allen-Bradley

Rockwell PLC

SAP / Dynamics

ERP integration

One data layer on top of all your machines.

The integrations above we have standardised across several industrial projects. For a new assignment, we set up sensor streams with the same abstraction layer: raw signals via OPC UA or MQTT to an edge broker, pre-processing in the cabinet, and aggregated features sent to the cloud for model training. That means fewer integration bugs, a faster lead time, and your own automation team can manage it once we step away.

For legacy machines without native connectivity, we build retrofit solutions: additional sensors, a gateway with edge compute, and a translation layer into your modern stack. This brownfield approach is exactly where off-the-shelf packages get stuck. Read more about our data engineering approach and computer vision applications for visual quality control.

At the ERP and CMMS layer we typically work with SAP, Microsoft Dynamics, Ultimo, Planon and Maximo. Alerts from the AI model are not delivered as an extra inbox stream but land directly as work orders in your existing system, with the right priority, a parts suggestion, and a link to the asset history. This means your maintenance planning, time registration and cost reporting keep working just as you're used to. For the technician on the shop floor, we provide a mobile app that fits seamlessly into that work order flow.

Finally, we are fundamentally opposed to vendor lock-in. The code, models, training data and infrastructure-as-code remain yours. On handover you receive full documentation, and your own team, or another party, can take over. We see that as a mark of quality, not a commercial threat.

From first audit to production-ready AI model.

An AI implementation on the shop floor has its own rhythm, different from a typical software project. Five phases that are identical for every manufacturing company.

01 · Audit

Shop-floor walk

A day on site with your maintenance manager and plant manager. Result: top three critical assets, available data, and a priority use case.

02 · Data foundation

Sensors & pipeline

Sensor selection, connectivity, edge broker, cloud bucket. Collect data first, then train the model. A few sprints to a working ingest.

03 · Model development

Anomaly & RUL

Anomaly detection on operational data, followed by RUL models where there is enough failure data. Validation alongside existing inspection rounds.

04 · Integration

CMMS & mobile

Translate alerts into work orders in your CMMS, a dashboard for operators, and a mobile app for the technician, without anyone needing a new login.

05 · MLOps

Ongoing

Models retrained on new data, drift detection, retraining pipeline. A fixed monthly management cycle, not a one-off delivery.

Work in manufacturing.

Food · Multi-line

Vibration monitoring on a packaging line

Edge sensors on critical motors, an anomaly detection model with CMMS work order integration. Operator dashboard on tablets at the line.

−42%

unplanned downtime

14 assets

monitored live

Machine building · OEM

Predictive platform for sold machines

The OEM supplies predictive maintenance as a service with new machines. Multi-tenant data platform, customer portal and SLA reporting per customer.

Service revenue

new revenue stream

Multi-tenant

one platform, multiple customers

Chemicals · Process plant

Yield optimisation for batch reactor

Tuning process parameters based on AI-driven recommendations, alongside existing SCADA. Operators stay in control, the model advises.

+3,1%

yield on pilot line

Operator-in-the-loop

no autonomous control

Recognised within the Dutch manufacturing industry.

FME Manufacturing Monitor 2024

"AI applications in predictive maintenance made their definitive breakthrough in Dutch manufacturing in 2024, particularly among mid-market producers who cannot carry the licensing structures of the large platforms."

Smart Industry Roadmap NL

"The greatest OEE gains lie in brownfield implementations: fitting existing machines with sensors and building an open data layer on top, rather than replacing the entire plant."

Customer assessment, food producer

"The combination of edge inference and a mobile app for the technician has genuinely changed how the line operates within the first few months: alerts reach the right person at the right moment."

Answers for the plant manager starting with AI.

The questions we most often hear from maintenance managers, plant managers and OEM product directors.

What exactly is predictive maintenance, and how does it differ from preventive maintenance?

Preventive maintenance follows a schedule (a service every 1,000 operating hours), regardless of the machine's actual condition. Predictive maintenance uses continuous sensor data and AI models to determine an asset's real health and to intervene only when a deviation emerges. It avoids both unnecessarily early maintenance and unplanned downtime. It is most effective on critical rotating machinery (motors, pumps, compressors, conveyor belts) with measurable degradation signals.

Which sensors do we need to get started?

It depends on the type of machine. For rotating assets, vibration (accelerometers) is the most valuable first sensor, followed by motor current, temperature and, for hydraulic systems, pressure. Sensors often already exist in the PLC, and we need not add new ones, only expose them via OPC UA or Modbus. We always begin a project with an audit to establish what data is already available before recommending any hardware investment.

How much data do we need for a working model?

For anomaly detection (deviation from normal behaviour), a few months of operational data is often enough to start. For genuine Remaining Useful Life prediction, you also need failure data: examples of assets that have actually failed. In a factory where breakdowns are rare, this takes longer. We are realistic about this and start with anomaly detection as the foundation while RUL data builds up.

Our machines date from the 1990s, is that really feasible?

Brownfield installations are exactly where our custom approach adds value. Older machines often have no native connectivity, but they can be monitored perfectly well with retrofit sensors and an edge gateway. We tap signals from the existing control system or fit external sensors to housings, bearings and motors. The big standard platforms expect greenfield – we don't.

When can we expect to see a return on investment from such a project?

That depends heavily on your starting position. In a factory with regular unplanned downtime and high downtime costs per hour, the business case often breaks even within several months of going live. For a plant that is already well maintained, the gains are smaller and lie more in labour efficiency and deferred replacement. We always carry out a baseline measurement beforehand (current OEE, average downtime costs) so that ROI can be assessed objectively afterwards.

Does our AI solution fall under the EU AI Act?

For pure predictive maintenance models, the classification is usually limited and not high-risk. Where AI touches safety functions (worker-safety camera detection, automatic machine stop), you are in high-risk territory with stricter requirements around data quality, transparency and human oversight. We map that classification from the scoping phase and document conformity. Also read our page on software for the manufacturing industry for the wider compliance context.

What does a predictive maintenance project cost?

A pilot on a single line or asset group with a handful of critical machines is a more compact project. A factory-wide platform across several production sites with deep ERP/CMMS integration is a larger project, depending on the number of assets, sensors, integrations and compliance requirements. We work with a fixed sprint budget and provide a concrete price for the full build after the audit and scoping phase.

Do you work with edge computing or do you do everything in the cloud?

For most predictive maintenance use cases, a hybrid approach is the right answer. Raw sensor streams (vibration at high sampling rates) don't belong in the cloud, so we pre-process them locally on an edge gateway in the control cabinet, where the first layer of anomaly detection runs. Only aggregated features and alerts move to the cloud for long-term history, model training and cross-asset correlation. For worker safety and computer vision, inference runs locally in any case, as latency and bandwidth don't allow a cloud round trip.

What if you step away, can we manage it ourselves?

Yes, that is a starting point. We deliver all code, training data, model artefacts and infrastructure-as-code into your own repositories and your own cloud account. We also provide a written architecture handbook and run at least one handover session with your own automation team. We offer management contracts, but never as a dependency: your team can always carry on independently.

Talk to us about your predictive maintenance project?

A thirty-minute first conversation with your maintenance or plant manager. We listen to your critical assets, ask questions about your data maturity, and give an initial direction. No obligation.

Book a shop-floor audit Email fabian.vandijk@appfront.nl