(528f) Explainable One-Class Classification Neural Network Model for Tablet Quality Control | AIChE

(528f) Explainable One-Class Classification Neural Network Model for Tablet Quality Control

Authors 

Taylor, A., MathWorks
Holt, R., MathWorks
Typical quality control tests applied to tablets include the content of the active ingredient or absolute drug content tests, uniformity of weight, uniformity of the content, disintegration time test, and dissolution tests. Using visual inspection systems, tablets are also inspected for cracks, chips, unusual colors, and other defects.

AI-driven methods are increasingly used in manufacturing processes for quality control as they promise less manual work and higher accuracies than human assessment. In this talk, we present an application of an explainable one-class neural network model for tablet quality control. With a one-class learning approach, accurate models could be developed even if the anomalies are scarce or change over time. Additionally, with this approach, smaller amounts of data are needed, and the labeling effort required is low – the knowledge of “good” and “bad” images is sufficient. The model is trained in MATLAB with images of normal pills and synthetically generated anomaly images. Following the recommended Fully Convolutional Data Description (FCDD) approach, a couple of realistic examples of anomaly images were added to the training data as outlier exposure to improve learning. The model produces a heatmap with the probability that each pixel is anomalous. The classifier labels images as normal or anomalous based on the mean value of the anomaly score heatmap. The model yields 97% accuracy in its classifications. We also demonstrate how similar AI models can be deployed to larger production systems.