(9a) Image Classification to Automate at-Line Pellet Inspection in Plastics Production

For many applications it is critical to deliver contamination free polyethylene product. Therefore, continuous inspection of polymer production is vital. Most systems inspect a pellet slip stream in an at-line fashion and provide feature information about any present contaminant such as color and size. However, in certain scenarios it is also of interest to determine whether the contamination is free flowing in the pellet stream (loose) or incorporated into the polymer pellet (embedded). Typical analytical equipment does not provide this information and the classification is thus a manual and subjective task.

In order to automate this classification, a multi-class classifier was built with a convolutional neural network (CNN). The network layers will automatically learn complicated features from raw images compared to traditional machine learning methods, which require well-defined features based on domain knowledge about the output class. A selection of well-established CNN model structures (i.e. VGG16, ResNet50, etc.) were tested with and without image augmentation and model accuracy evaluated. The best performing model achieves greater than 95% accuracy for all output classes. This is a successful application of deep learning techniques to directly improve manufacturing efficiency and the model is in use for daily decision making.