A general fault diagnosis framework for rotating machinery and its flexible application example

Abstract

When dealing with the fault diagnosis of different rotating machines (gear or bearing), different working conditions (such as rotating speed), different signals (acoustic signal or vibration signal), it is usually necessary to establish different models, which is, however, time-consuming and laborious. At the same time, the models have poor generality and portability. In order to solve above problems, a general fault diagnosis framework (GFDF) is proposed in this paper. Firstly, the collected signals, whether vibration signals or acoustic signals, are directly converted into two-dimensional gray images; secondly, the FAST-Enhanced-Unoriented-SIFT (FEUS) feature extraction algorithm proposed in this paper is used to extract feature vectors; then, the feature vectors are encoded by using the bag-of-words (BoW) model to obtain the basic words and codebook vectors; finally, the fault diagnosis is completed by calculating the distance between the description vector of the signal to be diagnosed and the codebook vectors. GFDF’s main feature lies in the evitable frequency domain transformation and noise reduction, which makes GFDF insensitive to signal type and has high diagnostic efficiency. The experimental results show that GFDF has high diagnostic accuracy and stability for acoustic signals and vibration signals of rolling bearing and planetary gear at different rotating speeds, which proves that GFDF has generality and portability and is potential for the application in other scenes. Comparative experiments show that GFDF outperforms the representative traditional classification methods and deep learning models in diagnostic accuracy and stability. In addition, GFDF is applied to the fault diagnosis of the acoustic signals collected in motion to simulate the working state of inspection robots, and the ideal diagnostic result is also achieved. The flexible application example of this framework provides experience for other researchers. © 2022 Elsevier Ltd