Imbalanced fault diagnosis: adversarial representation learning and category confusion for aero bearing

Abstract Existing unsupervised domain adaptation methods have predominantly been effective in scenarios where class balance is maintained. However, in the real-world operation of aircraft engines, the data generated from aero bearings is largely composed of normal data, with fault data being rare and difficult to collect, thus inevitably leading to class imbalance. Moreover, when domain shift and class imbalance occur concurrently, the performance of traditional methods tends to deteriorate significantly. To tackle this challenge, this paper proposes an imbalanced domain adaptation (IDA) method for aero bearings, based on adversarial representation learning and category confusion. Firstly, smooth domain adversarial training is introduced, which converges to the smooth minimum to effectively reduce the differences between domains. Secondly, a strategy of minimizing class confusion is proposed, aimed at enhancing the precision of pseudo-labels in the target domain. Additionally, a cost-sensitive learning module is developed through the use of penalty matrices, which compels the model to pay more attention to the underrepresented categories. Finally, extensive experiments are designed to verify the effectiveness and superiority of the proposed method.