Speaker
Description
The fusion reactor relies on multiple auxiliary equipment clusters such as vacuum, low temperature, and water cooling to work together. How to ensure effective monitoring and fault diagnosis of these devices is the key to the stable operation of the fusion experiment.Under different working conditions, there are obvious differences in the fault characteristics of rotating machinery, which directly leads to the failure of the model trained under a single working condition. Considering that there are often multiple working condition data sets in real scenarios, how to make full use of multi-source data sets to ensure model generalization performance becomes the key to fault diagnosis. Aiming at the problems of limited depth, lack of feature extraction ability and poor generalization ability of current multi-source models, an improved deep transfer learning algorithm called Two-stage Deep Fusion Attention Transfer Network (TS-DFATN), is studied and proposed. Firstly, the algorithm preprocesses the data sets of different working conditions, and obtains the image-based feature representation through time-frequency analysis. At the same time, a two-stage neural network was designed to extract features. The primary network was used to extract common features from different datasets, followed by multiple sub-networks to extract distinctive features from each dataset. In order to effectively extract differential features, a dual attention network was added to the subnetwork to enhance the capability of extracting local features. This method was tested on a public bearing dataset, and the results showed that the method improved the accuracy and generalization ability of the model.
| Minioral | Yes |
|---|---|
| IEEE Member | No |
| Are you a student? | No |