Improved Joint Distribution Adaptation for Fault Diagnosis

In the blast furnace (BF) ironmaking process, it is difficult to obtain labeled fault samples and the probability distribution drifts significantly. Therefore, transfer learning has been introduced to fault diagnosis of BF. Most of the existing transfer learning methods achieve domain adaptation by reducing the marginal and conditional distribution discrepancies, without considering the prior distribution. To address this issue, this paper provides a theoretical derivation of the effect of the prior distribution discrepancies on knowledge transfer, and proposes a new method called Improved Joint Distribution Adaptation (IJDA). The model performs reconstruction of the weighted source data to offset the discrepancies of prior distribution, and extracts domain-invariant features by aligning joint distribution to achieve knowledge transfer. In the transfer BF fault diagnosis experiments, the proposed method achieves promising performance improvement.