Data science consultancy in Eindhoven

It builds models using your own machine, sensor and maintenance data to predict outcomes such as a breakdown, a rejected part or the need to replace a component. These models are part of the AI applications we build and put into use.

If you’re first looking for advice on where to start with AI in your factory, see AI consultant Eindhoven. That’s about deciding what to do; here, we’re talking about the model itself.

No. Fokker Services is an aviation MRO company, providing maintenance, repair and overhaul services. We mention this case study as an example of what we do with maintenance and sensor data. We haven’t published a case study involving an organisation in Brainport.

Our office is on Westerdoksdijk in Amsterdam. In Eindhoven and the rest of Brainport, we work on site or online.

At Fokker, maintenance was largely scheduled at fixed intervals and based on engineers’ experience. This sometimes meant maintenance happened too early or too late. Working closely with the maintenance engineers, we translated their knowledge into features that make the model reliable.

A feature is an attribute the model uses in its calculations, such as the number of operating hours since the last overhaul. The engineer knows which attributes matter; the model weighs them. Read the case study.

Usually, it’s sensor readings from machine controls, downtime and faults recorded in an MES or maintenance system, orders and routings from the ERP system, and measurement reports from quality control. The challenge is linking them: which machine number and timestamp correspond to which fault?

Brainport Development describes the region as strong in high mix, low volume, high complexity manufacturing. That means there’s little repetition for each product. So we often make predictions per machine or component, rather than per product. Read more about the sector in data and AI for manufacturing.

There are then few examples to learn from, and a model that never predicts a breakdown will already be right almost all the time. That’s why we measure precision (how often an alert is justified) and recall (what proportion of breakdowns the model detects), then work with the planner to set the threshold.

If there are too few breakdowns, a model can instead learn what normal behaviour looks like and flag deviations. This is called anomaly detection.

For each prediction, we show which features carried the most weight, for example using SHAP values: a method that calculates each feature’s contribution to an individual prediction. The engineer can then see whether the model is responding to a rising temperature or to something irrelevant.

This also tests whether the features make sense. If a model performs well because of a feature that engineers consider illogical, it’s relying on a chance pattern in the data.

Where planners and technicians already work: in the planning system, maintenance system or a dashboard. At Fokker Services, the tool helps planners schedule work orders, parts and capacity ahead of time. To learn how we improve planning itself, see optimising planning with AI.

Machines, suppliers and products change. So agree in advance who will monitor the model and when it will be retrained.

When breakdowns are rare and inexpensive, when the manufacturer prescribes the maintenance schedule, or when a dashboard using ERP and MES data already answers the question. To learn how we build models for images, text and predictions, see machine learning.

For predictions about individual patients or customers, with an emphasis on the decision point and validation, read data science consultancy Utrecht.