Optimizing Data Training Quantity for Bearing Condition Monitoring

Abstract

Prognostics Health Management (PHM) in manu-facturing seeks to reduce the amount of unexpected downtime that inhibits manufacturing competitiveness. However, a common challenge for the manufacturing industry is the lack of known failure data to train a predictive classifier. This work optimizes the necessary quantity of required failure training data and healthy data for three different exemplar datasets by assessing classifier performance. Particle swarm optimization with penalty factors associated with the training data amount were used to identify the required training data amount for fault classification. Two separate analysis cases are considered: a binary classification and multi-class classification case termed the progressive case. In both analysis cases, the optimal training data depended on how separable the bearing data were between the different baseline and defect stages. In those instances where the differences in the data classes were apparent, the bearing data optimal training data amount was lower than in those instances where the data class differences were not present. Future work focuses on the investigation of these overlap cases to determine the best means for classifying progressive damage for remaining useful life calculations.