Stochastic Feed-forward Attention Mechanism for Reliable Defect Classification and Interpretation

Defect analysis in manufacturing systems has been crucial for reducing product defect rates and improving process management efficiency. Recently, deep learning algorithms have been widely used to extract significant features from intertwined and complicated manufacturing systems. However, typical deep learning algorithms are black-box models in which the prediction process is difficult to understand. In this study, we propose a stochastic feed-forward attention network that consists of input feature level attention. The stochastic feed-forward attention network allows us to interpret the model by identifying the input features, dominant for prediction. In addition, the proposed model uses variational inferences to yield uncertainty information, which is a measure of the reliability of the interpretations. We conducted experiments in the field of display electrostatic chuck fabrication process to demonstrate the effectiveness and usefulness of our method. The results confirmed that our proposed method performs better and can reflect important input features.