Bearing fault diagnosis based on capsule network and transfer learning

Bearings are widely used in many machines and equipment and often fail when they are used for long periods under different conditions of varying frequencies and loads. As a result, the detection of defects in bearings in various working states is crucial. However, the conventional neural networks are associated with severe feature extraction problems and relatively low detection rates due to data loss in the pooling layers. This study uses capsule neural networks to overcome the above - mentioned limitations in bearing fault diagnosis. This paper also solves the problem of traditional diagnostic methods that do not easily adapt to different conditions by using transfer learning with the capsule network. The experimental results are compared with the Case Western Reserve University’s dataset to check the efficiency of the proposed methodology for cross-condition bearing fault diagnosis. It is evidenced by the findings that this method is accurate, and that the algorithm introduced in this work is quite complex.