Robust bearing degradation assessment method based on improved CVA

Owing to the importance of rolling bearings in mechanical systems, many meaningful works have been presented for bearing degradation assessment. However, most of contributions fail to take into account the influence of input features on the performance of a degradation model. As a potential multivariable process control method considering serial correlations, canonical variate analysis (CVA) is introduced to detect abnormal behaviours of rolling bearings. This study combines the CVA and support vector data description (SVDD) to develop a robust approach to weaken or remove the negative impact of input data on bearing degradation estimation. First, the time-domain features extracted from healthy vibration signals are fused by CVA to generate projection data for SVDD training. Second, the distance of new variables to the centre of the hypersphere is calculated as an indicator to describe the degradation of rolling bearings. To demonstrate the effectiveness and robustness of the proposed method, two different experiments on bearing run-to-failure tests are conducted. Experimental results validate that the proposed work is capable of detecting bearing failure under various working conditions. Besides, comparison of the proposed work with the original CVA is implemented. The robustness of the proposed work is enhanced obviously compared with the original CVA.