A transition stage exists during the equipment degradation, which is between the normal condition and the failure condition. The transition stage presents small changes and may not cause significant function loss. However, the transition stage contains the degradation information of the equipment, which is beneficial for the condition classification and prediction in prognostics. The degradation based condition classification and prediction of rotating machinery are studied in this chapter. The normal, abnormal, and failure conditions are defined through anomaly determination of the transition stage. The condition classification methods are analyzed with the degradation conditions. Then the probability of failure occurrence is discussed in the transition stage. Finally, considering the degradation processes in rotating machinery, the condition classification and prediction are carried out with the field data.
TopBackground
Model-driven and data-driven methods are two important approaches taken by prognostics (Heng, Zhang, Tan, and Mathew, 2009; Byington, and Stoelting, 2004; Goebe, Saha, and& Saxena, 2008). Model-driven prognostics is established by the mathematical model of the physical component or statistical model of the certain failure mode. Consequently, model-driven prognostics presents higher accuracy but with a specific application range. Data-driven prognostics is implemented by analyzing the monitoring data as well as the history data (Schwabacher, and Goebel, 2007). Extensive adaptability characteristic makes the data-driven prognostics with lower accuracy as fault mechanism is not considered. Other prognostics methods are also studied in many cases where evolutionary prognostics is a promising approach to predict the equipment’s fault onset and indicate the possibility of the failure (Roemer, Byington, Kacprzynski, and Vachtsevanos, 2006), and it is data-based.
The conditions in condition monitoring are usually classified into normal, and various faulty types. Yang Yang (2005) used condition classification to study the healthy and faulty states in a small reciprocating compressor. One normal condition and four faulty conditions of the roller bearing were classified in (Jack, and Nandi, 2002). In fault diagnostics, 14 faulty types of the turbo pump were classified by Yuan, and Chu (2006), while normal states and abnormal states were not considered. Kinds of condition classification methods are studied including the linear and nonlinear classifiers. S. J. Sixon consider the five common classifiers: Euclidean distance to centroids, linear discriminant analysis, quadratic discriminant analysis, learning vector quantization and support vector machines (Dixon, and Brereton, 2009), with the results that the accuracy of the classifiers depends on the structure of the data set.