A transferable deep kernel principal component analysis framework for industrial fault diagnosis under variable working conditions

This paper investigates the domain adaptation problem in cross-domain fault diagnosis, with the goal of addressing this diagnostic challenge under varying operating conditions in industrial systems. A novel transferable deep kernel autoencoder (T-DKAE) framework is proposed, which incorporates the hierarchical orthogonal decoupling mechanism, dynamic fine-grained sub-domain alignment module, and adaptive uncertainty-aware optimization strategy. First, to separate the domain alignment space from the orthogonal classification space, a hierarchical orthogonal decoupling mechanism is proposed. Specifically, to achieve cross-domain distribution alignment, the framework employs a deep autoencoder, followed by the introduction of a Cayley transform to implement Kernel Principal Component Analysis (KPCA) as an orthogonal denoising mechanism, which extracts domain-invariant features while filtering out domain-specific redundant noise. Second, a dynamic fine-grained sub-domain alignment module is developed which employs a dynamic weighting scheme to adjust the alignment granularity according to the training progress, thereby effectively mitigating distribution discrepancy. Third, to dynamically balance the loss weights among reconstruction, alignment, and classification, an adaptive uncertainty-aware optimization strategy is introduced, which leverages homoscedastic uncertainty. This strategy avoids time-consuming parameter selection and ensures robust convergence. Finally, the effectiveness and superiority of the proposed method have been validated through experimental results on the three-phase flow dataset. Specifically, T-DKAE achieves an average diagnostic accuracy of 86.78% across eight cross-domain tasks, exceeding the best competing method by 1.2 percentage points. Even at the lowest signal-to-noise ratio (SNR) of -5 dB, T-DKAE maintains an average accuracy of 67.16%, demonstrating superior noise robustness.