This paper addresses the challenges of complex fault diagnosis in rotating machinery due to the increasing complexity and variable operating conditions (e.g., rotational speed, torque variations), particularly in situations requiring domain adaptation. A novel multi-output classification (MOC) framework is proposed for domain adaptation in partially labeled target datasets. Unlike conventional multi-class classification (MCC), the MOC framework simultaneously classifies the severity levels of complex faults. By applying multiple single-task and multi-task architectures (including shared trunk and interaction-based designs) to the MOC formulation, we perform complex fault diagnosis in partially labeled conditions. Specifically, we propose a novel interaction architecture, the Residual Neural Dimensionality Reducer (RNDR), which enables selective information sharing between diagnosis tasks and improves classification performance in complex fault scenarios. We incorporate frequency hierarchical normalization to enhance domain adaptation performance for motor vibration data. Using a motor-based test setup, we evaluate the implemented complex fault conditions under six domain adaptation scenarios. Experimental results demonstrate superior macro-F1 performance compared to baseline models, and single-fault comparisons demonstrate that the structural advantages of RNDR are more pronounced in complex fault settings. We also confirmed that frequency-layer normalization is more suitable for fault diagnosis tasks than existing methods. Finally, we analyzed RNDR and other models with increased number of parameters under various conditions and compared them with the pruned RNDR structure.
