Artificial Intelligence & Machine Learning

Where it pays off

In a production environment, machine learning earns its place mostly by seeing changes that are too slow or too multi-dimensional for a person to catch. A bearing’s vibration spectrum drifts over weeks; the quality of a part out of an injection mould depends on the combination of eight parameters. Both are things an eye looking at one gauge at a time misses.

The applications we work on most often:

• Predictive maintenance. Extracting pre-failure patterns from vibration, current signature and temperature data on rotating equipment — motors, bearings, pumps, gearboxes — so that planned maintenance follows the actual condition of the machine rather than a date on a calendar.
• Anomaly detection. Learning normal operation only, with no fault labels, and flagging deviation. In industry this is usually the approach that works, because the data everybody has in abundance is “everything is fine” data.
• Quality prediction. Inferring, in real time and from process parameters, a property that otherwise arrives from the lab at the end of the shift — what the literature calls a soft sensor.
• Demand and production forecasting. Separating seasonality and campaign effects from the order history to feed planning with something better than last year plus ten percent.
• Process optimisation. Deriving the trade-off between energy consumption, cycle time and scrap from data, then proposing parameter sets.

How we work

We start with a data assessment, not with model selection. The questions are: at what resolution are the signals sampled, are the timestamps aligned, can stoppages and shift changes be told apart in the record, and are failure events written in a maintenance log or kept in somebody’s head. The step is dull but it determines everything after it — misaligned timestamps let a model see the effect before the cause, and the result is a model that is perfect on paper and useless on site.

For data acquisition our preference is to work over OPC UA, using the control layer’s own tag dictionary. If a new measurement point is required we say so at the start: adding a sensor is both cheaper and more honest than forcing a model to guess.

In modelling we start simple. On tabular process data the gradient boosted tree family (XGBoost, LightGBM) usually outperforms deep learning, trains on far less data, and produces feature importances an engineer can actually interrogate. We move to deep learning when the problem genuinely calls for it — raw signals, images, or time series with long dependencies. Every experiment stays recorded in MLflow: which data version, which hyperparameters, which metric.

For deployment the criterion is model rollout that never stops the line. The model runs in a container or, converted to ONNX, on an edge device; if the inference service crashes the process is unaffected — only the prediction stream stops, and that shows up as an alert.

What makes industrial data different

Scarcity of labelled data. The central contradiction of predictive maintenance is that in a well-run plant failures are rare, so there is little to learn from. That is why we often build one-class or reconstruction-error approaches instead of classifiers.

Class imbalance. A model that calls a defect occurring once in a thousand parts “never present” scores 99.9% accuracy and is worthless. So accuracy is never reported on its own; precision, recall and false alarm rate go with it.

The real cost of a false alarm. If an operator walks to the machine three times for nothing, there is no fourth time. Threshold selection is an operational decision, not a statistical one, and it is made together with the maintenance team.

Concept drift. A model is not something you install and leave. Input distribution and prediction quality are measured continuously, and crossing a drift threshold puts retraining on the table.

How this connects to the other solution areas

A model is only as good as the layer beneath it. If the tag dictionary is inconsistent or the sampling rate is too coarse for the process, no algorithm compensates for that — which is why we run machine learning work as part of the same plan as data acquisition and analytics. Vision-based inspection is built together with machine vision and camera systems: if the lighting and optics are wrong, the model hunts for a defect where the camera never saw it. And when a prediction has to turn into an action, the industrial automation layer is what carries it.