Leveraging systems’ non-linearity to tackle the scarcity of data in the design of Intelligent Fault Diagnosis Systems

Intelligent Fault Diagnosis Systems (IFDS) often rely on Deep Transfer Learning (DTL) to identify structural damage. However, these models typically require large, labeled datasets, which are difficult to obtain for real-world industrial structures where faults are rare. The authors address this data scarcity by exploiting the inherent non-linearities found in mechanical joints and interfaces.

Instead of relying on traditional data augmentation, the researchers developed a non-parametric method that captures the Frequency Response Function (FRF) of a structure at multiple, controlled excitation levels. By arranging these FRFs into a 2D color map—where the y-axis represents excitation level and the x-axis represents frequency—they create image-like representations of the system's state. They further augment this dataset by permuting the collected FRF data, creating a larger pool of training samples without needing synthetic data generation models like GANs.

The researchers validated their approach using a railway pantograph, a structure known for its non-linear dry-friction behavior. By applying seven distinct excitation levels (ranging from 1 N to 13 N), they generated enough data to train pre-trained CNNs to distinguish between an undamaged state, a missing bolt connection, and a reduced damping scenario. The results demonstrate that the model can effectively classify these conditions despite the initial scarcity of experimental data, proving that excitation-level diversity can serve as a powerful proxy for data volume.

This work provides a practical pathway for deploying deep learning in structural health monitoring where data is inherently limited. By shifting the focus from generating synthetic data to intelligently sampling the physical system's non-linear response, the method reduces the computational burden and the need for complex physical models. It offers a robust, model-agnostic framework that can be integrated into existing experimental testing routines, such as modal analysis, to improve the reliability of fault detection in critical infrastructure.