Open-set domain adaptation for fault diagnosis using Gaussian auxiliary classifier and separation loss

Abstract Open-set domain adaptation (OSDA) is crucial in intelligent fault diagnosis, enabling knowledge transfer from labeled source domains to unlabeled target domains while detecting unknown faults. Existing OSDA methods often rely on softmax scores or fixed class boundaries, leading to overconfident predictions and misclassifications. To address this, this paper proposes GACS-OSDA, which combines a Gaussian auxiliary classifier with separation loss for accurate fault classification and unknown fault detection. GACS-OSDA models each known class as a Gaussian distribution in feature space, using a Gaussian negative log-likelihood loss to align features with their class distributions while handling uncertainty. To improve class separability, a separation loss enforces the distance between class means, creating distinct distributions. For unknown fault detection, dynamic thresholding adapts to each class’s uncertainty, identifying samples that significantly deviate from known classes. Experiments on two fault diagnosis datasets show that GACS-OSDA surpasses other OSDA methods in classification accuracy and unknown fault detection.