A New Deep Transfer Learning Network for Fault Diagnosis of Rotating Machine Under Variable Working Conditions

Machine learning is promising in vibration signal based fault diagnosis because of its full use of big data and nonlinearity extracting capability. However, in real-world application, the network trained by a vibration signal dataset will be applied to target signal datasets with different distributions, which can be triggered easily by rotating speed oscillation and load variation. Hence, based on transfer learning, some vibration signal-based methods which are robust to working conditions have been proposed to address this problem. Nevertheless, most of them need target datasets in network training, and the network should be trained whenever it meets a new vibration signal dataset. So we construct a three-stage deep fault diagnosis network utilizing adaptive batch normalization (AdaBN), which is highly efficient for free of target datasets in training and does not need repeated training in its application. In the first stage, the vibration signal samples are processed into more regular and discriminative frequency spectra. In the second stage, a fourlayer AdaBN based deep neural network (DNN) is pre-trained by stacked autoencoders (SAE) and then finely tuned only using the source dataset. In the final step, the trained network is modified to diagnose samples from the target dataset. Extensive experiments on a gearbox and a bearing dataset, and comparisons with some other fault diagnosis methods verify its effectiveness.