Class-level Alignment Domain Adaptation Method for Non-Destructive Testing of Transformer Faults

In practical industrial applications, the fault diagnosis method based on artificial intelligence often has the problem of insufficient accuracy because of the distribution deviation between different signal sources. To solve the above problems, this paper proposes a Class-level Alignment Domain Adaptation(CADA). In CADA, we introduce a distribution alignment module based on optimal transmission (OT) to eliminate inter-domain differences in data from different domains. At the same time, to make the data distribution and category alignment closer, we use the feature fusion strategy to combine the feature information and label information. Then realize the joint alignment of the data distribution and category by aligning the fusion features of the two domains. Further, we use the dual form of Wasserstein distance to measure distribution differences, thus reducing the high time cost of traditional OT-based methods to measure distribution differences. We conducted comparison experiments based on real voiceprint monitoring data to evaluate the effectiveness of the proposed method. The experimental results demonstrate the promising performance of our approach.