Tensor-Based Denoising on Multi-dimensional Diagnostic Signals of Rolling Bearing

To improve fault diagnosis efficiency, a multidimensional denoising approach based on tensor decomposition is developed for solving multidimensional signal filtering. The monitoring signals are decomposed via truncated high-order singular value decomposition (T-HOSVD) to obtain their factor matrices. With L-curve criterion, the appropriate truncation parameters in the tensor factorization are determined to denoise and reduce dimension of signals. Then, the performance of sequentially truncated HOSVD (ST-HOSVD) is verified to quantify the correlation between the dimension of signal and computation complexity. The proposed ST-HOSVD approach is then applied to reduce noise in torque, current and vibration signals collected on bearing test rig, respectively. Experimental results show that the performance of the T-HOSVD on signals denoising are poor with the dimension increasing. The ST-HOSVD approach can well remove noise and retain the working status features as much as possible. This tensor-based multidimensional signal filtering will be powerful tool for dealing with heterogeneous and multi-aspect data. The computation complexity of L-curve algorithm will increase sharply with the dimension of signal increasing while optimizing truncated parameters, but that of ST-HOSVD will not vary so much. When the dimension of the tensor model is not too high, the effectiveness of L-curve-T-HOSVD is higher. Otherwise, the computation cost of the ST-HOSVD is relatively lower. Therefore, the priority should be given to the ST-HOSVD for denoising the higher dimensional diagnostic data set about the rolling bearing.