Health indicator construction for roller bearing based on an unsupervised deep belief network with a novel sigmoid zero local minimum point model

Abstract

Extracting bearing degradation curves with good smoothness and monotonicity as a health indicator lays a solid foundation for predicting the bearing’s remaining useful life. Traditional bearing health indicator construction methods generally have the following problems: (1) they require manual experience, such as manual labeling of data is burdensome when the amount of collected data is large, for feature extraction, selection, and fusion with other indicators and models because the methods rely on substantial expert experience and signal-processing technology; (2) deep belief networks in deep learning require engineering experts with rich experience to label the data, and because the degradation state of a bearing is constantly changing, it is difficult to rely on manual experience to distinguish and label it accurately; (3) owing to the noise in the data collected during the study, the extracted health indicator curve shows obvious oscillation and poor smoothness. In response to the above problems, this study proposes a model based on an unsupervised deep belief network and a new sigmoid zero local minimum point to eliminate health indicator curve oscillation and improve monotonicity. The main idea is that a deep belief network without a label output layer is used to extract the preliminary health indicator curve directly from the original signal, whereas the sigmoid zero local minimum point uses the average value based on a sigmoid function to reduce the weight of the current health indicator value to eliminate concussion, and then it uses the zero and local minimum points to further improve the monotonicity of the extracted health indicator without parameters. Finally, the superiority of the model proposed in this study (deep belief network–sigmoid zero local minimum point) is verified through a comparison of multiple bearing datasets and other models.