| Abstract [eng] |
Estimation of the remaining useful life (RUL) of bearings is usually performed using various methods, including machine learning, entropybased, or other more or less classical methods. This study aims to use the ranks of the associated Hankel matrix. Algebraic techniques in predictive diagnostics are still less common, although some applications of Hankel matrices, for example, have received more attention recently [2, 3, 4]. The study further adds to the development of Hankel matrixbased applications in predictive diagnostics. Bearing acceleration measurements with respect to the vertical axis (obtained from the scientific experimentation platform PRONOSTIA [1] are analysed. A fixed sliding window is used, and the list of associated Hankel matrices is formed. As in our previous research, the computation of the singular values of the Hankel matrices is carried out by using the SVD decomposition. The optimal parameters for matrix size, threshold for singular value importance are determined previously. It must be noted that the pseudo-rank of the Hankel matrix is considered, since almost all real-world sequences are contaminated by noise. To reduce said noise, the fixed window of the moving average is used. In this way, pseudo-rank is employed as the property of a vibrational signal and a predictor of RUL. A new step in the study is the development of the prediction model to receive multidimensional input: a list of pseudo-ranks. Analysis of realworld diagnostic data shows that the correlation between pseudo-rank and remaining useful life could be seen when there was about 5 to 10 \% RUL left. Thus, a one-dimensional predictor is only suitable for analysing short RUL. [...]. |
