Prediction Reliability Assessment Based on Mahalanobis Distance and GRU in the Application of Bearing RUL Analysis

Abstract

Bearing is a key component of rotating machinery, and its remaining life prediction is crucial for equipment maintenance. As known that vibration signal has been widely used for data-driven methods, while those methods have been fall in the following two problems. The first issue is that how to construct a useful healthy index that reflects the bearing degradation trend factor. The other one is that signal samples are susceptible to noise interference and prediction models are easily affected by parameters, which will lead the prediction results are prone to discrete and unstable phenomena. There is a crucial need is that how to achieve a high prediction precision with reliable forecast results. Focusing on the appeal issue, this paper proposes a novel prediction reliability assessment based on mahalanobis distance and GRU, which contains three steps. First, considering the differences in sample distribution, mahalanobis distance as a health index is built from the multivariate statistical features. Then gate recurrent unit, as a time series prediction-sensitive deep learning approach, is used to predict the bearing remaining useful life. In the third step, the distribution of prediction results is obtained by multiple predictions, where Bootstrap method is employed to resample the samples for a sound datasets construction. Through experiments and comparison, the reliability of the proposed method has been verified with the prediction bias analysis. It can be foreseen that the reliability prediction assessment is not only applicable to bearings, but can also is used for remaining useful life prediction in other field of electronic equipment.