Fewer false alarms during extrusion
Whether it’s a tube, cable or profile: in extrusion, it is often the surface that determines whether a product is ‘good or bad’. Bubbles, streaks, micro-holes or contamination should therefore be detected as early as possible – whilst the material is travelling through the line at high speed. However, this is precisely what makes optical inspection so challenging: conventional camera systems frequently classify harmless variations as defects and reject good products, whilst genuine defects sometimes go undetected.
“The extrusion process is a dynamic production process, which means that even products manufactured without defects often look slightly different,” explains Peter Droege, co-founder and CEO of Maddox AI. Even slight variations in reflection or changes in surface wetting are enough to cause this. Added to this is the fact that the product often rotates slightly around its own axis during the process. Traditional, rule-based image processing, however, struggles with this because it usually works with fixed threshold values.
Four cameras, 360 degrees, up to 600 metres per minute
With Maddox Line, the Tübingen-based company aims to solve this problem. The in-line inspection system consists of the ‘Maddox Endless’ hardware, featuring four cameras that capture the product from all angles (i.e. at a 360-degree angle). It is designed for diameters of 1–100 millimetres and line speeds of up to 600 metres per minute. In addition to hoses and cables, the system also inspects profiles made from materials including PE, PVC, TPE and EPDM.
The evaluation is carried out by AI models that are trained on the specific production process. They are designed to learn which deviations occur as part of the normal process and which indicate production faults. According to the company, this requires at least 50 images showing genuine faults. If there are too few, the platform generates synthetic fault images. The training takes around two to eight weeks. Incidentally, workers on the production line can also operate the system without any programming knowledge.
From detection to prevention
The system displays detected defects in real time; a dashboard then analyses them across entire rolls, bundles or individual metres. Maddox AI also offers a process agent. “Many of our customers make extensive use of this to identify the causes of defects and scrap,” reports Droege. Users can ask it in natural language (i.e. the language people use to talk to one another) what defects are occurring or how quality is developing. If data from upstream processes is integrated, it should also be able to identify the causes. As an example, Droege cites a response such as: “The contact pressure at station 3 was too high – this leads to scratches on the surface.”
In practice: from the finger test to 100 per cent inspection
The system is used, amongst other things, for windscreen washer hoses, brake hoses, medical hoses, data cables and cable protection systems. A manufacturer of medical hoses, for example, illustrates the impact this switch can have. Previously, two employees per shift inspected the goods manually by feeling for bubbles with their fingers. There were five to ten customer complaints per year. Today, the system automatically inspects all five hose variants at a speed of 50 metres per minute. Maddox AI quotes the customer’s feedback on this: according to this, pseudo-rejects have fallen by 85 per cent, and there have been no further complaints since the installation. The manufacturer estimates the savings at between 150,000 and 170,000 euros per year.
We regularly manage to reduce pseudo-rejects by over 90 per cent.
Performance-based business model
The alignment of key performance indicators, installation, data collection and model development are free of charge for the customer. The customer only makes a purchase once the system achieves the agreed key performance indicators in normal operation; thereafter, an annual SaaS fee applies. The investment risk therefore lies with the provider itself. According to the company, the solution pays for itself within 6 to 18 months.
“We see immense growth potential for our product,” says Droege, referring to extrusion. This applies not only to continuous products: for films and sheets, the company offers Maddox Stream, its own system designed specifically for web materials.
Author: Ann-Kathrin Ullrich | K-Mag
Ann-Kathrin Ullrich enjoys explaining how new ideas from research, technology and industry are changing everyday life. Alongside her work for EuroShop-Mag, she focuses on plastics and rubber in K-Mag: from new materials and processing methods through to testing technology and issues relating to the circular economy.
