A neural network approach for remaining useful life prediction utilizing both failure and suspension histories

• Published in: Mechanical systems and signal processing Vol. 24; no. 5; pp. 1542 - 1555
• Main Authors: Tian, Zhigang, Wong, Lorna, Safaei, Nima
• Format: Journal Article
• Language: English
• Published: Kidlington Elsevier Ltd 01.07.2010; Elsevier
• Subjects: Artificial neural networks; Bearing; Condition monitoring; Errors; Exact sciences and technology; Failure; Fracture mechanics (crack, fatigue, damage...); Fundamental areas of phenomenology (including applications); Learning theory; Life prediction; Mathematical models; Measurement and testing methods; Neural networks; Optimization; Physics; Prediction; Remaining useful life; Solid mechanics; Structural and continuum mechanics; Suspension history; Suspension history; Remaining useful life; Prediction; Artificial neural networks; Bearing; Rupture; Bearing(mechanics); Durability; Neural network; Modeling; Pump; Signal processing; Monitoring
• ISSN: 0888-3270; 1096-1216
• DOI: 10.1016/j.ymssp.2009.11.005

Artificial neural network (ANN) methods have shown great promise in achieving more accurate equipment remaining useful life prediction. However, most reported ANN methods only utilize condition monitoring data from failure histories, and ignore data obtained from suspension histories in which equipments are taken out of service before they fail. Suspension history condition monitoring data contain useful information revealing the degradation of equipment, and will help to achieve more accurate remaining useful life prediction if properly used, particularly when there are very limited failure histories, which is the case in many applications. In this paper, we develop an ANN approach utilizing both failure and suspension condition monitoring histories. The ANN model uses age and condition monitoring data as inputs and the life percentage as output. For each suspension history, the optimal predicted life is determined which can minimize the validation mean square error in the training process using the suspension history and the failure histories. Then the ANN is trained using the failure histories and all the suspension histories with the obtained optimal predicted life values, and the trained ANN can be used for remaining useful life prediction of other equipments. The key idea behind this approach is that the underlying relationship between the inputs and output of ANN is the same for all failure and suspension histories, and thus the optimal life for a suspension history is the one resulting in the lowest ANN validation error. The proposed approach is validated using vibration monitoring data collected from pump bearings in the field.