Quality Prediction in Injection Molding

TL;DR

This paper compares neural network architectures with classical regression methods for predicting injection molded part quality using inline measurements, showing neural networks outperform traditional algorithms despite limited data.

Contribution

It demonstrates the superior prediction performance of neural networks over classical regression methods in industrial injection molding quality prediction.

Findings

Neural networks outperform classical regression algorithms.

CNNs trained on thermographic images yield better predictions.

LSTM networks trained on raw signals also perform well.

Abstract

Injection molded part quality can be improved by precise process adjustment, which could rely on in-situ measurements of part quality. Geometrical and appearance quality (visually and sensory) requirements are increasing. However, direct measurement is often not feasible industrially. Therefore, process control must rely on a prediction of parts quality attributes. This study compares prediction performances of diverse neural networks architectures with "classical" regression algorithms. Dataset comes from inline industrial measurements. Regression was performed on 97 scalar statistical features extracted from multiple acquisitions sources: thermographic images and analog signals. Haralick features were extracted. Convolutional Neural Networks were trained on thermographic images and Long Short Term Memory networks were trained on raw signals. Although the dataset was small, neural networks show better predictions scores than other regression algorithms.