Fault Diagnosis of Rotating Machinery Using Trace Constraint Stacked Autoencoder

Stacked autoencoder (SAE) plays a significant role in dimensional reduction and feature selection; although conventional SAE has excellent performance in non-linear feature mapping, it is difficult to extract and select sensitive features suitable for fault diagnosis tasks of rotating machinery. To address this issue, a trace constraint stacked autoencoder (TC-SAE) is proposed by introducing label information into the whole training process of SAE. Feature extraction and fault diagnosis are accomplished using a improved SAE, which is able to obtain high classification accuracy. The availability of the proposed method is validated on both the rolling bearing data set of Case Western Reserve University and the data set generated through the equipment in our lab.