For Industrial Fault Diagnosis With Domain and Category Inconsistencies, a Multisource-Refined Transfer Network

PROJECT TITLE :

Multisource-Refined Transfer Network for Industrial Fault Diagnosis Under Domain and Category Inconsistencies

ABSTRACT:

In recent years, there has been a concentrated effort put into researching unsupervised cross-domain fault diagnosis. It does not require target supervision to learn transferable features that reduce distribution inconsistency between the source domain and the target domain. The majority of the currently available methods for cross-domain fault diagnosis are developed on the basis of the assumption that the source and target fault category sets are consistent with one another. This assumption, on the other hand, is generally put to the test in practice due to the fact that various working conditions can have different fault category sets. In this article, a multisource-refined transfer network is proposed as a solution to the problem of fault diagnosis under conditions where both domain and category inconsistencies are present. First, an adversarial multisource-domain adaptation strategy is developed with the goal of lowering the refined categorywise distribution inconsistency that exists within each source–target domain pair. It steers clear of the negative transfer trap that conventional global-domainwise-forced alignments are known to produce. After that, a multiple classifier complementation module is developed by transferring the source classifiers to the target domain and complementing them there. This is done so that different diagnostic knowledge that is present in different sources can be leveraged. The similarity scores generated by the adaptation module are used to supplement the various classifiers, and the complemented smooth predictions are utilized to direct the refined adaptation. Thus, the refined adversarial adaptation and the classifier complementation can benefit from each other in the training stage, yielding target-faults-discriminative and domain-refined-indistinguishable feature representations. Extensive testing on two different scenarios demonstrates that the proposed method is superior in situations in which both domain and category inconsistencies are present.

Did you like this research project?

To get this research project Guidelines, Training and Code... Click Here