Deep Domain Adaptation for Cross-Condition Fault Diagnosis in Hydraulic Systems under Data Scarcity

Data-driven fault diagnosis methods have achieved remarkable success in monitoring the health of hydraulic systems under constant operating conditions. However, in real-world industrial scenarios, hydraulic systems frequently operate under variable and dynamic conditions, such as fluctuating accumulator pressures, varying temperatures, or shifting external payloads. This variability inevitably leads to a severe discrepancy in the data distribution between the training (source) and testing (target) domains—a phenomenon widely known as domain shift. Such distributional discrepancies dramatically degrade the diagnostic performance and generalization capabilities of conventional deep learning models. To address this profound challenge and the critical issue of label scarcity in newly emerging working conditions, this paper proposes a novel Deep Domain Adaptation framework tailored for cross-condition fault diagnosis in complex hydraulic systems. Specifically, a multi-channel One-Dimensional Convolutional Neural Network is meticulously designed to extract hierarchical, multi-scale temporal representations from multi-sensor time-series data. Furthermore, the Maximum Mean Discrepancy with a multi-kernel Gaussian approach is seamlessly integrated into the loss optimization function. This explicitly minimizes the marginal distribution divergence between the label-rich source domain and the completely unlabeled target domain within a deep latent reproducing kernel Hilbert space. Extensive empirical experiments are conducted on the globally recognized UCI condition monitoring of hydraulic systems dataset. The comprehensive results demonstrate that the proposed method effectively decouples intrinsic fault characteristics from operational pressure variations, achieving an outstanding diagnostic accuracy of 97.5% in a severe cross-condition task (transitioning from 130 bar to 90 bar). Rigorous comparisons with several state-of-the-art baseline models, detailed ablation studies, alongside t-SNE feature visualizations and confusion matrix analyses, robustly prove the superiority, resilience, and industrial edge-deployment potential of the proposed framework.