Few-shot transfer learning with attention for intelligent fault diagnosis of bearing

The bearing is one of the key components in modern industrial equipment. In the past few years, many studies have been carried out on bearing diagnosis through data-driven methods. However, there are two practical problems. First, under actual working conditions, the lack of fault samples is a major factor that hinders the application of these methods in industrial environments. Second, there is a lack of full utilization of a priori knowledge in the current stage of methods using relational networks for fault diagnosis. It is manifested by the incompleteness of the relational network structure. To address these problems, we present a new diagnosis method based on few-shot learning, which is suitable for the environment where the data is scarce. In this method, we train the model with the data generated by the artificial damaged bearings instead of the data from the real bearing. We experimentally validate the performance improvement of the complete relational network structure. It is able to perform the few-shot learning task better. In addition, we also reduce the global feature discrepancy by introducing an attention mechanism to improve the performance of the model. And the impact of the number of layers of the attention mechanism on the model is also discussed in detail. In this paper, our model performs better under the same experimental conditions compared with other transfer learning models.