Data Augmentation for Intelligent Manufacturing with Generative Adversarial Framework

The global economy is greatly shaped by the unprecedented booming of ICT and artificial intelligence technologies. Their applications in manufacturing has led to the advent of intelligent manufacturing and industry 4.0. Data has become a precious asset for modern industry. This paper first introduces an energy monitoring and data acquisition system namely the Point Energy Technology, which has been developed by the team and installed in several industrial partners, including a local bakery. The lack of data always exists due to various reasons, such as measurement or transmission errors at data collection and transmission stage, leading to the loss of varied length of data samples that are key for process monitoring and control. To solve this problem, we introduce a generative adversarial framework which is based on a game theory for data augmentation. This framework consists of two multilayer perceptron networks, namely generator and discriminator. An improved framework with Q-net that extracts the latent variables from real data is also proposed, in which the Q-net shares the structure with discriminator except for the last layer. In addition, the two optimization methods, namely mini-batch gradient descent and adaptive moment estimation are adopted to tune the parameters. To evaluate the performance of these algorithms, energy consumption data collected from a bakery process is used in the experiment. The experimental results confirm that the latent generative adversarial framework with adaptive moment estimation could generate good quality data samples to compensate the random loss of samples in time series data.