A deformable CNN-DLSTM based transfer learning method for fault diagnosis of rolling bearing under multiple working conditions

Abstract

Machine learning methods are widely used for rolling bearing fault diagnosis. Most of them are based on a basic assumption that training and testing data are adequate and follow the same distribution. However, for bearings working under multiple working conditions, dynamic changes are inevitable and labelled vibration data are usually insufficient. To deal with the issues, a new fault diagnosis method using deformable convolutional neural network (CNN), deep long short-term memory (DLSTM) and transfer learning strategies is designed. Specifically, a model is constructed by integrating deformable CNN, DLSTM and dense layers. Among them, deformable CNN enhances the ability of standard CNNs for local feature extraction using fixed geometric structures. DLSTM further encodes the sequential information contained in the output of deformable CNN. Dense layers are applied to capture high-level features then classify the data samples as each fault type. The model is firstly pre-trained using data samples under one working condition. Then, transfer learning strategies are implemented to fine-tune the pre-trained model utilising very few samples of another working condition, enabling it to identify fault types of bearing under new condition. Experiments are conducted and results show that the presented model yields higher than comparative performance compared with state-of-the-art methods.