Fault diagnosis of overhead hoist transports (OHTs) is crucial in semiconductor manufacturing, where OHT failures can halt wafer transfers between tightly synchronized processes, leading to significant downtime and potential wafer damage. However, developing a practically applicable fault diagnosis framework for a fleet of OHTs is challenging due to significant variability in torque signals across different units, the limited availability of labeled data, and the need for interpretability to support to support on-site decision-making. To address these issues, this article proposes a novel approach called the fleet-level semi-supervised domain adaptation network, which enables robust and interpretable OHT fault diagnosis. The proposed method employs a semi-supervised domain adaptation strategy to mitigate domain discrepancies between units and enhance diagnostic performance using both unlabeled and labeled data. Also, the method processes dual-motor torque signals from the front and rear motors to physically meaningful signals and extracts features using a multi-head convolutional neural network (CNN) structure. A feature-weighting module is incorporated to dynamically highlight informative features, which not only enhances diagnostic performance but also improves the interpretability of the diagnostic process. The validation of this method is performed using a dataset logged from OHT units that were in actual operation across multiple semiconductor manufacturing lines, demonstrating superior fault diagnosis performance and high practical applicability under limited labeling conditions. Moreover, the model provides interpretable diagnostic insights by analyzing multi-head weight contributions, enabling a more reliable assessment of its health conditions.