Ball Screw Health Indicator Construction with Limited Monitoring Data and Health Assessment Based on Global Context Network

Abstract

Ball screw is one of the most important transmission mechanisms in manufacturing. Constructing effective health indicator to represent the health condition of ball screw and assessing health state are of great significance to reducing unnecessary cost. However, researches on ball screw health indicator and health assessment are limited. Most health indicator construction methods for rotating machine require manual feature extraction process and feature selection or reduction process using monitoring data at different health states, which are difficult to obtain for ball screw in real applications. In this paper, a deep learning-based ball screw health indicator construction method with limited monitoring data and a health assessment method are proposed. The health indicator construction model is trained using partially monitored sensor data based on denoising convolutional autoencoder (DCAE) and maximum mean discrepancy (MMD). DCAE is used to extract feature representations from monitoring signals. Then feature distribution discrepancy measured by MMD is mapped to HI. An end-to-end health state assessment method based on global context network (GCNet) is also proposed. Experiments show the proposed method achieved better assessment accuracy.