Intelligent Fault Diagnosis with Domain Adaptation Techniques

Intelligent fault diagnosis harnesses data-driven methods and machine learning to detect, localise and classify defects in mechanical systems. A persistent challenge is the discrepancy between data collected under laboratory conditions and data encountered in diverse industrial environments, commonly referred to as domain shift. Domain adaptation techniques seek to bridge this gap by aligning feature distributions from a well-labelled source domain to an unlabelled or sparsely labelled target domain. Approaches range from closed-set adaptation, which assumes identical fault categories, to open-set and partial adaptation, where the target may exhibit unknown or fewer fault types. Generative models and adversarial training have become instrumental in learning domain-invariant representations and disentangling domain-specific features. These advances enable robust diagnostics across varying operating conditions, reduce negative transfer and improve the detection of novel fault classes. The integration of adaptive weighting schemes, pseudo-labelling and synthetic data generation has proven essential for enhancing transferability in safety-critical applications such as rotating machinery, gearboxes and turbines.

Research from Nature Portfolio

No recent Nature Portfolio content available.

Intelligent Fault Diagnosis with Domain Adaptation Techniques publication trend

The graph below shows the total number of articles in intelligent fault diagnosis with domain adaptation techniques across all publications each year (not limited to Nature Index journals).

Technical terms

Domain adaptation: A set of methods for transferring knowledge from a labelled source domain to an unlabelled or differently distributed target domain.

Domain shift: The difference in data distribution between training (source) and deployment (target) environments that degrades model performance.

Partial domain adaptation: A scenario in which the source domain contains more fault categories than the target, requiring selective transfer to avoid negative transfer from irrelevant classes.

Open set assumption: A setting where the target domain may include fault classes that were not present or labelled in the source domain, necessitating methods to recognise unknown categories.

Adversarial network: A model comprising competing components (generator and discriminator) that learn to produce feature representations indistinguishable across domains to achieve domain-invariance.

References

1. Domain adaptation with domain specific information and feature disentanglement for bearing fault diagnosis. Measurement Science and Technology (2024).
2. Controlled generation of unseen faults for Partial and Open-Partial domain adaptation. Reliability Engineering & System Safety (2023).
3. Cross-Domain Open Set Fault Diagnosis Based on Weighted Domain Adaptation with Double Classifiers. Sensors (2023).