A Novel Semi-Supervised Deep Transfer Learning Method With Improved Extended Isolation Forest for Bearing Fault Diagnosis

Industrial condition monitoring leverages transfer learning to enhance equipment diagnostics. However, existing deep transfer learning (DTL) methods face a critical challenge, which still requires substantial annotated samples for robust fault diagnosis, and training deep models remains time-consuming and labor-intensive. To address this gap, we propose a semi-supervised deep transfer method based on an improved extended isolation forest (EIF) for cross-domain bearing fault diagnosis. First, nontarget task samples are scored using an EIF with an arithmetic mean aggregation mechanism for anomaly scores to identify latent fault patterns. A hybrid metric integrating the Kolmogorov–Smirnov (K–S) test and confidence-driven Hellinger distance is employed to generate pseudolabels for anomaly score samples. Subsequently, deep learning (DL) models are trained on these labeled data. Finally, a minimal quantity of target-domain labeled data is used to refine the pretrained models to complete cross-domain bearing fault diagnosis. Extensive validation tests on the HUST bearing dataset and a self-constructed dataset demonstrate that the proposed method achieves high diagnostic accuracy with significantly fewer labeled samples. The experimental results demonstrate that the proposed novel approach with a fine-tuning strategy achieves over 20.52% higher diagnostic accuracy than traditional direct transfer approaches on both the HUST bearing dataset and the self-constructed dataset while reducing the required labeled samples for fine-tuning to only 5%.