Context
Part quality is checked after ejection, away from the machine. Measuring every shot is not practical, so sampling is the norm and whatever was molded between two checks goes unrecorded. Moisture in the granulate, mold temperature, a change of material lot: all of these move the quality, and the first sign of it is usually a part already out of tolerance.
The machine, meanwhile, records injection pressure and injection flow every few milliseconds. Those signals are not the quality itself, but they carry what happened during the shot that produced the part.
What we tested
We built a model that estimates part quality from the signals the molding machine already records. Nothing was added to the machine. The model is built on the family of lightweight machine-learning techniques we work with across our projects (reservoir computing).
Validation used a public dataset in which real parts were molded on a production machine while process conditions were varied deliberately, with the quality of every shot measured. The targets were part weight, a dimension, and straightness. Accuracy was measured on an interval used neither for training nor for tuning.
What it shows
Weight, dimension and warpage could all be estimated on the held-out interval. The figure is available a few seconds into the cycle, so every shot carries an indication of quality without waiting for the part to reach the measuring bench.
The model this takes fits in a few hundred bytes and runs on the existing controller or PLC as it is. No GPU, no cloud connection.
On a process condition the model has not seen, the estimate is normally off. Measuring a small number of parts from that condition and adding them to the training brought it back to the accuracy reached with the full set of data for that condition; there is no need to rebuild the dataset from scratch every time a condition changes.
This may apply to your process
If the parts between two checks carry no quality figure, or a condition change costs you time before the process settles, the same idea may be worth exploring. The signals it needs are ones your machine already records. Which signals matter, and how they have to be handled, differs from process to process and decides whether it works at all. A short conversation is normally enough to tell whether your case fits.
Inferring quality from a waveform is not specific to injection molding: pressing, welding, heat treatment, coating and filling lines share the same construction. The continuous-process version is described on the quality soft sensor page.