Generating Information-Diffusion-Based Virtual Samples to Improve Small Data Set Prediction for Ceramic Powder: A Case Study

A multi-layer ceramic capacitor is a widely used passive component in modern electronics. However, most passive component manufacturers have to undertake pilot runs after receiving a batch of the key component, ceramic powder, to confirm the dielectric constant because of its low stability among batches. It takes at least two weeks from the pilot runs to mass production. In order to reduce the costs, one effective way is to predict the dielectric constant with experiential data. Although neural networks are widely applied to implement this task, the learning models are usually built with a great amount of training data, which is difficult to collect in the early stages of a manufacturing system. Therefore, this paper on the basis of the information-diffusion concept generates more training samples to help improve the prediction. The results reveal that the proposed method can rapidly help develop a model of production with limited data.