Contextual Bandit Guided Data Farming for Deep Neural Networks in Manufacturing Industrial Internet

Deep Neural Networks (DNNs) have shown superior performance in supervised learning in Industrial Internet applications, such as quality modeling, virtual inspection, etc. However, the performance of DNNs relies on large and high quality data sets with sufficient sample sizes and appropriate distributions. Additionally, collecting and labeling large data sets can be labor-intensive and may fall short of meeting the online computational needs, including the fact that more samples in training may not always improve the modeling performance. Inspired by the theory of Design of Experiments, we propose a Contextual Bandit-based Representation Design (CBRD) to generate data suitable for training DNNs. The CBRD combines the offline experimental design criteria as the arms of a contextual bandit model for DNN training in a joint and interactive way for online batch data. A low-dimensional representation of the input design space learned by variational autoencoder is used to generate new samples. The integration of VAE and contextual bandit enables the generation of samples adaptive to modeling performance. A real case study of Aerosol® Jet Printing process is used to demonstrate the merits of the CBRD method.