Virtual Reinforcement Learning for Defect Prediction in Smart Manufacturing

Recent research has focused on the integration of smart manufacturing and deep learning owing to the widespread application of neural computation. For deep learning, how to construct the architecture of a neural network is a critical issue. Especially on defect prediction or detection, a proper neural architecture could effectively extract features from the given manufacturing data to accomplish the targeted task. In this paper, we introduce a Virtual Space concept to effectively shrink the search space of potential neural network structures, with the aim of downgrading the computation complexity for learning and accuracy derivation. In addition, a novel reinforcement learning model, namely, Virtual Proximal Policy Optimization (Virtu-PPO), is developed to efficiently and effectively discover the optimal neural network structure. We also propose an optimization strategy to enhance the searching process of neural architecture for defect prediction. In addition, the proposed model is applied on several real-world manufacturing datasets to show the performance and practicability of defect prediction.