Co-training partial domain adaptation networks for industrial Fault Diagnosis

TL;DR

A novel PDA framework called Interactive Residual Domain Adaptation Networks (IRDAN), which introduces domain-wise models for each domain to provide a new perspective for the PDA challenge, and establishes a reliable stopping criterion for selecting the best-performing model.

Abstract

The partial domain adaptation (PDA) challenge is a prevalent issue in industrial fault diagnosis. Drawing inspiration from traditional classification settings where such partial challenge is not a concern, we propose a novel PDA framework called Interactive Residual Domain Adaptation Networks (IRDAN), which introduces domain-wise models for each domain to provide a new perspective for the PDA challenge. Each domain-wise model is equipped with a residual domain adaptation (RDA) block to mitigate the ADP problem. Additionally, we introduce a confident information flow via an interactive learning strategy, training the modules of IRDAN sequentially to avoid cross-interference. We also establish a reliable stopping criterion for selecting the best-performing model, ensuring practical usability in real-world applications. Experiments have demonstrated the superior performance of the proposed IRDAN.

Figures (3)

• Figure 1: The structure of IRDAN.
• Figure 2: The interactive learning strategy for IRDAN. The left figure shows the forward (solid lines) and back (dashed lines) propagation caused by samples from source (blue) and target (red) domains and the black dashed lines represent the MMDs. The right figure shows the general step of the interactive training from step 1 to step 6.
• Figure 3: the correlation between the accuracy (blue lines) and the reward (red lines) for the proposed IRDAN and DANN. The four subfigures on the top show the accuracy and reward of IRDAN at each epoch on tasks B5, B9, B11 and B12 for the Paderborn dataset while the subfigures on the bottom represent the accuracy and reward of DANN for the same tasks. Triangles on the red lines note the maxima of the reward and the circles on the blue lines show the selected models based on the reward. Triangles on the blue lines show the real best accuracy.