TY - JOUR
T1 - A Novel Sensor-Based Label-Smoothing Technique for Machine State Degradation
AU - Chao, Ko Chieh
AU - Shih, Yu
AU - Lee, Ching Hung
N1 - Publisher Copyright:
© 2001-2012 IEEE.
PY - 2023/5/15
Y1 - 2023/5/15
N2 - Due to the rise of Industry 4.0, machinery health prognosis has become one of the primary objectives of machine maintenance. Long-term operating machines gradually age over time, resulting in machine degradation and lower production yields. In general, the entire life cycle of a machine is from health to degradation to fault states. Once a machine breaks down, it may increase production costs and cause serious safety hazards. To prevent the machine from operating in a fault state, a variety of sensors are applied for machine health monitoring. Subsequently, the collected sensor data are fed into a degradation model, which is used to evaluate the machine degradation level. As the machine degradation process changes continuously over time, the features in the transition region between two adjacent condition states are nearly identical. Similar features lead to a poor degradation model performance in the transition region. In this study, a novel label-smoothing method is proposed to improve model performance in the transition region. To enhance the machine degradation model accuracy, the proposed method consists of four major processes: preprocessing, data label smoothing, model training, and postprocessing. The developed model is verified using a bearing run-to-failure dataset and a tool wear dataset, achieving a prediction accuracy of over 90%. The experimental results demonstrate that the proposed method outperforms other existing peer methods, especially in the transition region.
AB - Due to the rise of Industry 4.0, machinery health prognosis has become one of the primary objectives of machine maintenance. Long-term operating machines gradually age over time, resulting in machine degradation and lower production yields. In general, the entire life cycle of a machine is from health to degradation to fault states. Once a machine breaks down, it may increase production costs and cause serious safety hazards. To prevent the machine from operating in a fault state, a variety of sensors are applied for machine health monitoring. Subsequently, the collected sensor data are fed into a degradation model, which is used to evaluate the machine degradation level. As the machine degradation process changes continuously over time, the features in the transition region between two adjacent condition states are nearly identical. Similar features lead to a poor degradation model performance in the transition region. In this study, a novel label-smoothing method is proposed to improve model performance in the transition region. To enhance the machine degradation model accuracy, the proposed method consists of four major processes: preprocessing, data label smoothing, model training, and postprocessing. The developed model is verified using a bearing run-to-failure dataset and a tool wear dataset, achieving a prediction accuracy of over 90%. The experimental results demonstrate that the proposed method outperforms other existing peer methods, especially in the transition region.
KW - Bearing health monitoring
KW - convolution neural network
KW - label smoothing
KW - machine degradation
KW - sensor selection
KW - tool wear
UR - http://www.scopus.com/inward/record.url?scp=85153499534&partnerID=8YFLogxK
U2 - 10.1109/JSEN.2023.3263634
DO - 10.1109/JSEN.2023.3263634
M3 - Article
AN - SCOPUS:85153499534
SN - 1530-437X
VL - 23
SP - 10879
EP - 10888
JO - IEEE Sensors Journal
JF - IEEE Sensors Journal
IS - 10
ER -