BRENNEN

Industries / Mining and materials / Plastics and rubber

Quality checks

Quality work makes sure shaped plastic and rubber parts come out right: testing the material, watching the process while parts are made, measuring and inspecting finished parts, and running the written systems that customers audit.

6

Milestones achieved, from 1924 to 2026.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1924

    Control charts read variation

    A control chart plots a measurement over time against limits, so a worker can tell normal variation from a change that needs action. The first record of Walter Shewhart's control chart is an internal memo dated 16 May 1924, which Quality Digest describes as a Bell Telephone Laboratories memo. The chart is the root of the statistical process control (SPC) that molders use today.

    qualitydigest.com

    Credit

    Walter A. Shewhart

  2. 1987

    One rulebook for quality

    ISO published ISO 9001, 9002 and 9003 as models for quality management systems, building on British standard BS 5750. A quality management system is the written set of procedures, records and audits a company uses to keep its output consistent. The certification body DQS counts more than 1.2 million certified organizations.

    dqsglobal.com

    Credit

    ISO Technical Committee 176; British Standards Institution

  3. 1999

    Carmakers add their own rules

    The International Automotive Task Force (IATF), formed by carmakers and their trade associations, created ISO/TS 16949 in 1999 to add car-industry requirements on top of ISO 9001. Its successor, IATF 16949:2016, "cancels and replaces ISO/TS 16949:2009".

    iatfglobaloversight.org

    Credit

    IATF

  4. 2000

    Sensors watch every shot

    A cavity pressure sensor sits inside the mold and measures the pressure of the plastic as it fills the part, so a bad shot can be flagged and rejected. RJG released its eDART monitoring system in 2000, and the medical molder Comar says eDART is "installed on every injection-molding machine" it runs. RJG says its founder built the first in-mold cavity pressure sensor in 1962, which is RJG's own claim.

    rjginc.compcinternational.com

    Credit

    Rod Groleau and RJG; Kistler also makes such systems

  5. 2018

    CT scans see inside parts

    Industrial computed tomography (CT) takes X-ray images from many angles and builds a 3D model of the part, so inspectors can measure hidden features without cutting it open. By 2018 Quality Magazine called CT "ideal for the first article inspection of injection molded plastic components". Plastics processors such as Pexco now offer it for first-part checks, shrinkage and warpage.

    qualitymag.compexco.com

    Credit

    CT makers such as Zeiss; users such as Pexco

  6. 2026

    US rules adopt ISO 13485

    From 2 February 2026, the FDA's rules for how medical devices are made, renamed the Quality Management System Regulation, "incorporate by reference" ISO 13485:2016, the international quality standard for medical devices. Molders of medical parts now work to the ISO standard under US law.

    fda.gov

    Credit

    US FDA; ISO

Watch for

  1. AI grades parts from pressure

    Kistler says its ComoNeoPREDICT software "uses artificial intelligence to calculate the quality of each individual part" from cavity pressure and temperature signals in the mold, so every part could be checked without being measured. It gives no accuracy figure and no outside test.

    Achieved whena medical or automotive molder publishes how well the predictions matched real measurements on a large sample, including bad parts missed and good parts rejected.

    kistler.com

    Claimed by

    Kistler, September 2025

  2. AI camera checks each part

    In a demonstration cell, a camera captures each molded part, AI judges its quality, and bad parts are sorted out. Beckhoff says "only a small number of good and bad parts" were needed to train the model, and gives no accuracy figure.

    Achieved whena production user publishes measured detection and false-reject rates against human or gauge inspection.

    beckhoff.com

    Claimed by

    Nissei Plastic and its subsidiary Negri Bossi, with Beckhoff, September 2025 (K 2025)

Open

  1. 01

    Defect photos across part designs

    A public, labeled photo set of molded parts showing sink marks, short shots, flash, burn marks, weld lines (visible seams where two melt fronts meet) and good parts, from at least three different part designs. Add a baseline detection model tested on a part design it never saw in training.

    The brief
    Credit · unclaimed
  2. 02

    Two-machine part quality benchmark

    A public dataset and benchmark that tests whether a model trained on one injection molding machine can predict part weight and one key dimension on a second machine running the same mold and material. Trained staff at a partner molder or technical college run the machines, and the builder handles only the data and the weighing.

    The brief
    Credit · unclaimed

Open 01 · the brief

Defect photos across part designs

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public, labeled photo set of molded parts showing sink marks, short shots, flash, burn marks, weld lines (visible seams where two melt fronts meet) and good parts, from at least three different part designs. Add a baseline detection model tested on a part design it never saw in training.

Why it is still open

An academic set of 2,500 images has no sink mark or short shot class, and its page says "Files have not been uploaded for this dataset". The dataset in Roboflow's June 2026 tutorial uses other classes: break, crack, defect, pollution and poor. The SKZ and Fraunhofer IPA public dataset labels defects on one part design, a "warpage shell", and its download lists tables, not image files.

ieee-dataport.orgroboflow.comb2share.eudat.eu

Built with what exists

Scrap parts donated by local molders, so no machine work is needed; a Raspberry Pi High Quality Camera or any camera with fixed lighting; CVAT, an open-source labeling tool; RF-DETR, an open detection model under Apache 2.0, as the baseline; Zenodo or Hugging Face to host the data under CC BY 4.0.

  1. First step

    Ask two local molders for a box of scrap parts already sorted by defect type, and photograph 100 of them under one lighting setup.

  2. Done looks like

    At least 2,000 labeled images, with at least 200 per defect class. The test split uses a part design held out from training. Precision and recall are published for each class: how many flagged parts are truly bad, and how many bad parts are caught.

Open 02 · the brief

Two-machine part quality benchmark

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public dataset and benchmark that tests whether a model trained on one injection molding machine can predict part weight and one key dimension on a second machine running the same mold and material. Trained staff at a partner molder or technical college run the machines, and the builder handles only the data and the weighing.

Why it is still open

The public sets found come from a single machine, such as the SKZ and Fraunhofer IPA set of 303 parts from one KraussMaffei 160-750PX, labeled OK or not OK. A 2024 cross-machine study had to use Moldex3D simulation to stand in for its second machine.

b2share.eudat.euscholar.lib.ntnu.edu.tw

Built with what exists

Shot-by-shot machine data through the EUROMAP 77 (OPC UA) interface or the machine's own export; a lab balance reading to 0.001 g; calipers or an optical gauge; Python with scikit-learn or LightGBM; the SKZ and Fraunhofer IPA set (CC BY 4.0) for a first baseline.

  1. First step

    Download the SKZ and Fraunhofer IPA dataset and reproduce an OK/not-OK baseline, and write to two molders or colleges to ask for a shot-data partnership.

  2. Done looks like

    At least 1,000 shots per machine, with weight-prediction error on machine B published for two cases: a model trained only on machine A, and one trained on machine A plus 50 shots from machine B. Data and code are released under an open license.