Dynamic Gated Reverse Capsule Network for fault diagnosis of rolling bearings

Aiming to address the challenges of low fault diagnosis accuracy and poor generalisation performance caused by variable operating conditions and insufficient fault samples, a rolling bearings fault diagnosis method based on Dynamic Gated Reverse Capsule Network (DGRCN) is proposed. First, the Markov Transition Field (MTF) method is adopted to transform the original one-dimensional vibration signal into a two-dimensional feature image with temporal correlation. Second, leveraging the low computational complexity of depthwise separable convolution and the advantage of dilated convolution to expand receptive fields without increasing parameters, a Dynamic Gated Separable Dilated Convolutional Network (DG-SDCN) module is designed by dynamically adjusting dilated convolution weights using a gated network, which optimises feature extraction while enhancing the model’s expressiveness and flexibility. Subsequently, to ensure high efficiency while preserving rich spatial information and feature hierarchy, the DG-SDCN is further embedded into a Reverse Capsule Pruning Network (RCPN) to construct a Dynamic Gated Reverse Capsule network to improve fault diagnosis performance. Finally, the proposed model is validated on two distinct datasets. Experimental results demonstrate that the method achieves superior classification accuracy, generalisation capability and robustness under varying loads, speeds and dataset scales.