A Novel Label Smoothing Technique for Machine Degradation

Abstract

Due to the rise of Industry 4.0, most factories use fully automated equipment to reduce labor costs and increase production efficiency. These long-running machines gradually age over time, which results in machine degradation and lower product yields. In general, the whole life cycle of a machine is from health state to degradation state to fault state. Once a machine breaks down, it may increase production costs and leads to serious safety hazards. To prevent the machine from running in a fault state, many sensors are applied to monitor the health of the machine. Then, the collected data are fed into the degradation model, which is used to evaluate the degradation level. Because machine degradation is a continuous process, the features in the transition region between adjacent two condition states are nearly identical. Similar features make the degradation model perform poorly in the transition region. In this study, a novel label smoothing method is proposed to improve the model performance in the transition region. The proposed method which is tested on a bearing run-to-failure data dataset has achieved a prediction accuracy of 96.76%. The results of the experimental study demonstrate that the proposed method outperforms the other compared peer methods.