AI on the Factory Floor

The extrusion of insulated cables is a high-volume, high-precision process. Every metre produced outside specification translates into wasted copper and insulation material, rework, or scrap. Quality assurance depended on operators watching the line and deciding when to start “good” production, a task that is both repetitive and critical.

Manual quality checks on the line

‍Quality verification relied on human intervention. Operators monitored the extrusion process and decided when the cable had reached the right dimensions and properties. This increased the risk of error and made outcomes heavily dependent on individual experience and attention.

High material waste during start-up

Delays in decision-making and conservative safety margins meant unnecessary losses of insulation material at the start of each run. Every second spent producing out-of-spec cable translated into direct cost.

Variable standards across recipes

The start-up phase differed by recipe and cable design, making it harder to ensure consistent product quality. With hundreds of recipes and variations, standardising behaviour was difficult without a systematic way to detect when the process had stabilised.

The work focused on automating the manual quality assurance step and optimising insulation material usage during extrusion, without requiring new hardware or major infrastructure upgrades.

A machine learning classifier, trained on historical production data, runs directly on the factory floor machines. In real time, it determines when the cable is approaching target properties and signals the moment when “good” production can begin. This replaces the operator’s visual judgement with a consistent, data-driven decision.

Inference time was a critical constraint. Each second without a decision meant costly material waste, so predictions had to be effectively instantaneous. The model was designed to deliver real-time outputs using only the minimum data required, ensuring that decisions keep up with the line speed.

On top of the classifier, a controller optimises raw material usage by adjusting parameters in real time. A digital twin methodology was used to simulate and refine controller behaviour before deployment, reducing risk and tuning performance without disrupting production.

The entire solution was packaged as a lightweight Python application with minimal dependencies, compatible with legacy factory systems.

Each machine now saves over €300 per day in raw material, with internal measurements indicating savings closer to €350 per day in some lines.

More than 1000 extrusion recipes are modelled, covering a wide range of cable designs and operating conditions.

Around 12 metres of copper and insulation material are saved on every run by shortening the out-of-spec start-up phase.

By automating quality verification, operators are freed from constant monitoring and can focus on higher-value tasks, such as overseeing multiple lines or handling exceptions.