A method of automatic feature extraction from massive vibration signals of machines

In the studies of intelligent fault diagnosis of machines, lots of effort goes into designing effective feature extraction algorithms. Such processes would consume plenty of human labor, especially when dealing with massive vibration signals. So it is interesting to automatically extract features using machine learning techniques, instead of manually extracting them. To deal with the problem, this paper presents a new automatic feature extraction method of machines. The proposed method first learns features from the vibration signals by K-means, and then maps the learned features into a salient low-dimensional feature space using t-distributed stochastic neighbor embedding (t-SNE). Through the feature extraction results of a bearing dataset, it is verified that the proposed method is able to effectively learn the features from the raw vibration signals and is superior to the manual features like time-domain features and wavelet features. Therefore, the proposed method has potential to be a tool in the automatic data mining of intelligent fault diagnosis.