Senvis-Net: Learning from imbalanced machinery data by transferring visual element detectors

Abstract

With the development of sensor network technologies and the popularization of Industry 4.0, data-driven machine health monitoring has become increasingly important, not only to save maintenance costs of factory machinery, but also to guarantee the safety of factories. However, traditional data-driven algorithms are limited by two aspects. Firstly, the information fusion of multiple sensors heavily relies on domain knowledge. Secondly, imbalanced distribution of machinery data brings a challenge for the machine learning algorithm performance. In order to tackle these issues, we propose a general methodology to organize collected sensor data into image form and utilize visual element detectors learned by a pre-trained convnet to explore meaningful information hidden in data. We also design a Convolutional Neural Network (CNN) model, named Senvis-Net. Applied to an imbalance learning task of remaining useful life (RUL) prediction, our model outperforms the state-of-the-art CNN that learns directly from sensor data. Moreover, transferring visual element detectors can bring another 20.1% ~ 97% performance benefits depending on severity of imbalance.