Semi-supervised Fault Identification Based on Laplacian Eigenmap and Deep Belief Networks

Abstract

Aiming at solving the problem of insufficient labeled samples in mechanical equipment fault diagnosis, a semi-supervised fault recognition model based on Laplace Eigenmap (LE) and Deep Belief Network (DBN) is presented by combining the idea of manifold learning and deep learning. The model utilizes LE algorithm to directly extract the features from the raw high-dimensional vibration signal, and inputs the low-dimensional manifold features into DBN. By using a few expensive labeled samples and lots of cheap unlabeled samples, it excavates fault features for a second time, and finally constructs the Soft-max classification to identify the fault mode of the mechanical equipment. The semi-supervised model is applied to the identification of bearing faults and gear cracks. The test results show that LE algorithm can effectively reduce the time complexity of the model, enhance the intelligence of feature extraction and improve the diagnostic efficiency. DBN network can fully mine fault characteristics, get better feature representation, and improve diagnostic accuracy. In addition, the model also achieves a good diagnosis effect with unbalanced training label, and is applicable to the diagnosis of multi-sensor feature fusion, which has practical application value.