Leveraging systems' non-linearity to tackle the... | AI Research

Leveraging systems' non-linearity to tackle the scarcity of data in the design of Intelligent Fault Diagnosis Systems
This paper introduces a new method for diagnosing faults in mechanical structures when very little data is available. Intelligent Fault Diagnosis Systems (IFDS) typically rely on Deep Transfer Learning (DTL), which usually requires massive amounts of labeled data to function effectively. Because real-world machine faults are often rare or difficult to replicate in a lab, the authors developed a technique that uses the natural non-linear behavior of physical structures to create more data from limited experimental tests.

Using Non-Linearity to Create Data

The core idea is that many real-world structures, such as those with joints or friction, behave differently depending on the strength of the force applied to them. By applying vibrations at several different excitation levels, the researchers can capture how a structure’s frequency response changes under different conditions. They convert these frequency response measurements into 2D color maps, where the vertical axis represents the excitation level and the horizontal axis represents frequency. This creates a visual "spectrogram" of the structure's health.

A New Augmentation Technique

To address the lack of data, the authors propose a custom data augmentation method. By taking the matrices of frequency response functions collected at different excitation levels and swapping specific rows, they can generate a large number of new, realistic color maps without needing to perform additional physical experiments. This approach allows them to leverage the power of pre-trained Convolutional Neural Networks (CNNs)—which are typically used for image recognition—to classify the health of the structure. Unlike other methods, this technique does not require training a complex generative model from scratch, making it computationally efficient.

Experimental Validation

The researchers tested their method on a real-scale railway pantograph, a structure known for its complex, non-linear dry-friction joints. They simulated two specific damage scenarios: the loss of a bolted connection and the removal of an artificial damper. By applying seven different levels of excitation, they were able to generate enough data to successfully train an IFDS to distinguish between the undamaged structure and the two faulty conditions.

Key Considerations

The proposed method is non-parametric, meaning it extracts features directly from raw data without needing complex physical models. It is particularly well-suited for structural dynamic tests where an external actuator can control the amplitude of the vibration. While the method effectively overcomes data scarcity, it relies on the fundamental assumption that the structure exhibits non-linear behavior that changes in response to varying excitation levels. This makes it a specialized but powerful tool for structural health monitoring in industrial settings where data is hard to come by.