Anomaly Cause Classification in the Extrusion Process Using Prompt-Based Segmentation and Convolutional Neural Networks

In this study, a method was developed to detect anomalies caused by incorrect process condition settings in the extrusion process for catheter tube manufacturing and to classify the causes of these anomalies. The proposed method first employs a prompt-based segmentation technique to segment the extruded resin, and then extracts the time history of the area and height ratio of the segmented resin to classify the cause of the anomaly. A specially developed deep learning network is designed to use the signal in the time domain to classify the cause of the anomaly, while also visualizing and analyzing the reasoning in the frequency domain. This network performs feature extraction on the time history of the height ratio and area, using a softsign function and power pooling to preserve frequency domain information as much as possible, thus enabling visualization of the basis for determining the cause of the anomaly. We collected a total of 47 minutes and 53 seconds of data under 14 process conditions and organized it into seven classes for analysis. The proposed method achieved 100% classification accuracy using data obtained under the given process conditions, demonstrating nearly perfect results.