关键词: 特征提取, 故障诊断, 降噪自编码器, 深度学习

Abstract: An asymmetric frequency domain auto-encoder method was proposed to extract spectrum characteristics of vibration signals. Unlike traditional auto-encoder, which took reconstructed vibration signals as the target output, the frequency domain auto-encoder used the frequency spectrum of vibration signals as the target output. This asymmetric auto-encoder might learn the mapping between vibration signals and the frequency spectrum, so that the encoder might output frequency domain features. In order to illustrate the feature extraction effectiveness of the proposed frequency domain auto-encoder, feature extraction and fault diagnosis experiments were carried out on bearing data sets. Without introducing label information, the features extracted by the frequency domain auto-encoder show better clustering effectiveness and may distinguish different fault types of the bearings. In the further generalization experiment, using 1% labeled samples in training classifier may achieve more than 90% of fault classification accuracy. The experimental results show that the frequency domain auto-encoder may extract fault feature information from vibration signals better than that of the traditional auto-encoder and has certain practical value.

Key words: feature extraction, fault diagnosis, denoising auto-encoder, deep learning