A Deep Learning Method for Rolling Bearing Fault Diagnosis through Heterogeneous Data

Abstract

Vibration signals of rolling bearing have multiple heterogeneous forms. Traditional fault diagnosis methods use 1D time-series signals or converted 2D signals for fault diagnosis. However, using the former will lose the spatial neighborhood features; using the latter will ignore time-series features, which caused information waste. In this paper, a new heterogeneous form of bearing vibration signals is proposed to address the problem. Our contributions of include: First, we proposed dynamic waveform sequences, which is a new heterogeneous form and can simultaneously reflect time-series features and spatial neighborhood features in vibration signals. Second, the CCLSTM (Conv-ConvLSTM) model is designed to extract the above two features layer by layer. Relying on the powerful feature extraction capability of CCLSTM, it is possible to simultaneously extract the time-series features and spatial neighborhood features in a single fault diagnosis network. The experimental verification through real bearing fault data sets shows that this method can effectively improve the diagnostic accuracy.