Remaining Useful Life Prediction Of Rolling Elements Bearings Based On Multiple Health State Assessment

Bearings are one of the most common key components in mechanical machines and prediction of the remaining useful life (RUL) of bearings can help engineers take timely actions to prolong the life of equipment and avoid failures and economic losses. This thesis provides an approach for RUL prediction of rolling bearings based on multi-health state assessment.This thesis first introduces the basic mathematical principles of SVM. Kernel functions are used to map sample data from the original space to the Hilbert space in SVM models. SVC and SVR are two aspects of SVM applications. LIBSVM is the toolbox used in this paper to process relevant calculations.The PRONOSTIA platform is then introduced to illustrate accelerated life testing of bearings. The original data collected from sensors must be preprocessed before being used for RUL prediction. We firstly smooth the original data before feature extraction to lower the interference of noises. Features used in this paper include time domain features, frequency domain fatures, entropy features, and EMD features. The total number of features is 88, including 44 in the horizontal direction and 44 in the vertical direction.Through analyzing the extracted features, we find that the Hilbert entropy can divide the bearing life process into two conditions:life including state Ⅰ, state Ⅱ and state Ⅲ, or life including only state Ⅰ and state Ⅲ. Before assessing the health state, appropriate features must be selected from feature database. Correlation analysis is used to calculate the correlation coefficients between features and health states. The selected features are the input of the support vector machine classifiers (SVC) and the corresponding labels are the output of SVC. Five-fold cross validation and grid search, which is used twice, is used to optimize parameters of SVC. The obtained results show that the optimized SVC can provide good health assessment.Bearings in state Ⅰ remain in a good health condtion and the feature indexes are in a steady level. Thus, we don’t predict RUL in this stage. Bearings begin to degrade in state Ⅱ and the features selected by correlation analysis gradually change. State Ⅲ means that the bearings degrade sharply and they can fail at any time. So we build separate SVR models for state Ⅱ and state Ⅲ. The validation results show that both SVR models can provide good RUL prediction through comparing prediction life with reported research results.The whole bearings life cycle can be divided into different health states through SVC models, and different SVR models are built to predict the RUL of bearings. This approach can provide relatively accurate prediction results. The last part of this thesis provides a summary of the work done and points out its shortcomings. Future research directions are provided as well.