An Unsupervised Method for Rolling Bearing Fault Diagnosis Based on Clustering and Stacked Auto-Encoder

In traditional supervised methods, model training requires a large number of manual labeled samples, which are time-consuming and labor-intensive. A new unsupervised method for rolling bearing fault diagnosis based on clustering and stacked auto-encoder is proposed in this paper. The new method provides a high-efficiency model, which consists of three parts. Raw vibration signals are extracted into multidimensional feature matrix, of which the dimension is further reduced by laplacian eigenmaps (LE) algorithm. The reduced feature matrix are then clustered using a K-means algorithm to form the training set, which are the inputs into a stacked auto-encoder (SAE). By fine tuning the parameters of the deep network, the trained SAE is employed to classify fault type and fault size. The experiment uses the proposed method to perform fault diagnosis from the data of rolling bearing system, and the results show that the method is able to distinguish different fault categories and sizes accurately and effectively. More importantly, it indicates that the proposed method removes the initial steps of artificial calibration samples and thus improves the intelligent level.