Non-local Denoising Convolutional Neural Network for Rolling Bearing Vibration Signal

摘要

The denoising process for rolling bearing vibration signal could significantly influence the performance of the diagnosis, prognostics and control process of a mechanical system. During the last several years, convolutional neural network(CNN) has been widely applied in the field of image denoising. However, it has been rarely applied to time-series signal denoising. Therefore, in this paper, we construct a novel architecture of CNN for rolling bearing vibration signal denoising. Furthermore, we also adopt non-local means (NLM) which is firstly introduced in the field of image denoising to construct non-local block (NLB), which could improve the performance of our proposed CNN considerably. Moreover, the NLB could be plugged into the denoising architecture at any positions without breaking the original behavior of the network. Based on it, we also propose a non-local denoising convolutional neural network (NL-DeCNN) aiming at rolling bearing vibration signal denoising. In the experiment section, we firstly proposed our original CNN architecture and then test the influence of the single NLB when it is embedded into different layers. Next, we test the influence of the number of NLBs. Lastly, we also compare our proposed CNN based method with other three traditional time-series denoising methods and our NL-DeCNN exhibits the best performance among these methods.