A deep-learning model with improved capsule networks and LSTM filters for bearing fault diagnosis

Abstract

The deep-learning networks for bearing fault diagnosis may encounter ambient noise interference. The inherently feedforward serial structure lacks the suppression of the ambient noise involved in bearing data. Moreover, the accuracy of the diagnosis model relies on training a large amount of labeled data, thus resulting in a significant amount of time consumption. To combat these two challenges, an improved capsule network diagnosis model with a long short-term memory filter (LF-iCapsNet) is proposed. First, the LSTM network is used to filter out the noise interference in the time domain through a nonlinear moving-average mechanism. At the same time, a subsequent feature-extraction convolution with large kernels is employed to specifically suppress the noise in the frequency domain, where a dilated convolution is adopted to achieve multi-scale feature extraction. Second, to obtain the distance-dependent relations of low-level features, an inner dependence operator is introduced into the primary capsule. Thus, a richer and more complete feature description of the bearing fault is guaranteed. And so, the derived digital capsule gives a geometrical constraint representation among the pixels in the feature maps and a rather high-speed training process because of its one-stage detection mode. Finally, the proposed LF-iCapsNet model is validated on the dataset from Case Western Reserve University (CWRU). The experimental results show that the diagnosis model provides a considerable improvement in the classification accuracy of bearing faults under the noise condition.