Support evidence statistics for operation reliability assessment using running state information and its application to rolling bearing

• The concept and algorithm of support evidence statistics are presented.
• A feature selection method using vibration signals is improved.
• A new experiment of rolling bearing is conducted to support the presented method.
• Another experiment downloaded from the PCE is employed to validate the conclusion.

Traditional reliability evaluation method generally requires a large amount of previous data and information on historical lifetime. For an individual mechanical device without historical lifetime data, it is difficult to carry out the reliability assessment by using the traditional method. To attempt exploring this difficult problem, support evidence statistics (SES) as an approach to operation reliability assessment is presented in this paper. Moreover, this presented method is also expected to indicate the physical state changes of the individual mechanical device. Since in scientific research, evidence usually goes towards supporting or rejecting a hypothesis. For a running device, evidences derived from the running state information should consistently demonstrate its current sole-running-state within a given short time interval. In practice, due to the interference of environmental noises, these evidences lose the consistency. Accordingly, they can be classified into two classes: the firm evidences and the flimsy evidences. Analogous to the support vector data description (SVDD), these firm evidences which show remarkable consistency can form a support evidence space (SESP) through one-class classification. Suppose that a SESP is obtained by using the evidences accumulated from the normal running state, the device operation reliability at any time of unknown running state can be evaluated through the statistical comparison between the normal SESP and the unknown SESP. This reliability evaluation process is named as SES. The most fundamental distinction between the proposed method and the traditional method lies in different statistical objects. The traditional methods are to analyze lifetime data while the proposed methods are to analyze running state data. Obviously, the evidence feature optimization plays a crucial role in the presented method. The maximum correlation and minimum redundancy (MCMR) method is improved by principal component analysis (PCA) to select evidence features based on vibration signals. Finally, the effectiveness of the presented method is validated through a new experiment of rolling bearing.