Intelligent fault diagnosis for low-speed roller bearings based on stacked auto-encoder

  • Hai-hong Tang
  • Zhi-qiang Liao Mie University, 1577 Kurimamachiya-cho Tsu-shi Mie-ken, Tsu-shi and 514-8507, Japan
  • Yusuke Kobayashi Railway Technical Research Institute Materials Technology Division Applied Superconductivity Laboratory, 2-8-38 Hikari-cho Kokubunji-shi, Tokyo and 185- 8540, Japan
  • Tomita Masaru Railway Technical Research Institute Materials Technology Division Applied Superconductivity Laboratory, 2-8-38 Hikari-cho Kokubunji-shi, Tokyo and 185- 8540, Japan
  • Peng Chen Railway Technical Research Institute Materials Technology Division Applied Superconductivity Laboratory, 2-8-38 Hikari-cho Kokubunji-shi, Tokyo and 185- 8540, Japan
Keywords: low-speed roller bearing, stacked auto-encoder, fault diagnosis, feature extraction

Abstract

There is an apparent issue that the bearing vibration signals in low-speed rotating are contaminated with strong noise. It is difficult to extract fault feature and effectively diagnose fault by traditional methods. For solving this problem, an intelligent fault diagnosis method based on stacked auto-encoder (SAE) is proposed to analyse low-speed roller bearing signal, which can effectively capture the representative fault information in signal with strong noise and achieve the advantages of deep architecture-based feature representations. The proposed method includes three successive parts. Firstly, the raw signal is pre- processed through the FFT and divided into training and testing sets for the SAE model. Secondly, the deep hierarchical structure is then established with the rule of greedy training, where the auto-encoders are utilized to obtain high-order characteristics of the training sets with weak fault feature. Thirdly, the testing sets
are applied to confirm bearing fault detection of low-speed roller bearing. The proposed method have been verified by low-speed bearing fault signal (10~70/100 RPM), the results show that the proposed method can achieve the high accuracy. Moreover, the comparative experiments with the BPNN and SVM are employed to prove the effectiveness of SAE that is prior to deal with the big data with strong noise and can overcome the dependent with manual selection and profession knowledge. The above experimental results adequately show that the SAE can adaptively extract the fault features in the raw signal with weak feature due to the strong noise and obtain high fault identification accuracy.

Author Biography

Hai-hong Tang

Mie University, 1577 Kurimamachiya-cho Tsu-shi Mie-ken, Tsu-shi and 514-8507, Japan

Published
2019-11-18
Section
Articles