Deep Learning Systems with Linguistic Interpretability in Manufacturing Image Classifications

Machine learning has received considerable attention in recent decades in data-driven modelling systems and methods. Machine Learning focuses on applied maths and computing algorithms for creating ‘computational machines’ that can learn to imitate system behaviours automatically. Unlike traditional system modelling methods (physics-based, numerical etc.), machine learning does not require a dynamic process model but sufficient data, including input and output data of a specific system. It thus could get high prediction accuracy but lack interpretability. A method to add transparency to deep CNNs is adding a fuzzy logic radius basis function to specific CNN structures named RBF-CNN. With the deletion of the defuzzy layer, a more generalised form was introduced, namely ND-RBF-CNN. Both RBF-CNN and ND-RBF-CNN were benchmarked for linguistic fuzzy rules' prediction accuracy and interpretability. Both structures demonstrated good interpretability at a small cost of prediction accuracy. To improve the prediction accuracy, a general RBF layer initialisation methodology was explored.