DGCDN: robust acoustic fault diagnosis via domain-generalized causal disentanglement

Abstract Reliable industrial fault diagnosis depends on the robust interpretation of sensor measurements, which are often degraded by unseen operating conditions. Acoustic sensing, a flexible and non-contact modality, is highly desirable for this purpose; however, its practical application is severely hindered by low signal fidelity and pervasive background noise. These measurement challenges introduce significant domain shifts that current domain generalization (DG) methods struggle to overcome, often failing to disentangle faint physical fault signatures from nuisance factors inherent in the measurement environment. To address these measurement-centric limitations, we propose the domain-generalized causal disentanglement network (DGCDN). Grounded in a structural causal model of the signal generation and measurement process, DGCDN learns domain-invariant health representations by causally isolating them from domain-specific measurement artifacts. This disentanglement is enforced through a synergistic multi-objective function that preserves measurement fidelity via reconstruction, ensures feature independence through an orthogonality constraint, and promotes domain invariance using contrastive learning. The training process is further stabilized by an adaptive domain weighting mechanism and entropy regularization, which respectively prioritize challenging source domains and mitigate predictive uncertainty. Validation on datasets exhibiting substantial shifts in measurement conditions confirms the model’s efficacy, as it consistently achieves state-of-the-art performance across a suite of evaluation metrics, including accuracy, precision, recall, F 1-score, and AUC. These results highlight DGCDN’s exceptional robustness to domain shifts and data scarcity, underscoring its potential for reliable deployment in high-stakes industrial measurement systems.