Bearing Fault Diagnosis: Class-Conditional Deep Domain Adaptation for Generalization Across Machines

Reliable fault diagnosis is essential for preventing unexpected failures and ensuring the operational safety of rotating machinery. However, most existing approaches assume identical data distributions and suffer significant performance degradation under varying operating conditions. To address this limitation, this paper proposes a domain-adaptive framework for generalized bearing fault diagnosis across heterogeneous environments. The method first transforms raw vibration signals into time-frequency scalograms using Continuous Wavelet Transform (CWT), enabling effective characterization of nonstationary and transient fault patterns. A deep convolutional feature extractor is then employed to learn discriminative representations, while a class-conditional alignment mechanism based on Local Maximum Mean Discrepancy (LMMD) is introduced to reduce intra-class distribution discrepancies between source and target domains. Furthermore, a pseudo-label driven subdomain adaptation strategy is developed to iteratively refine target domain alignment without requiring labeled data, thereby improving robustness in unseen operating conditions. The proposed framework jointly optimizes classification and alignment objectives to learn domain-invariant yet highly discriminative features. Extensive experiments conducted on two independent bearing datasets under different operating conditions demonstrate that the proposed method consistently outperforms conventional deep learning and existing domain adaptation approaches in cross-domain scenarios. Feature visualization further reveals improved intra-class compactness and interclass separability, confirming effective knowledge transfer. These results highlight the practical applicability of the proposed approach for real-world condition monitoring, where labeled data in new environments are often unavailable.