Overview
This document introduces a novel method for developing vibration-based Intelligent Fault Diagnosis Systems (IFDS) under conditions of limited data availability. The approach utilizes Deep Transfer Learning (DTL) and addresses the challenge of insufficient labeled data, which is common in diagnosing faults in machines and structures. It specifically leverages the intrinsic non-linearities of real-world systems through a periodic multi-excitation level procedure.
Research Context
Deep Transfer Learning (DTL) enables the efficient construction of Intelligent Fault Diagnosis Systems (IFDS). However, DTL methods typically necessitate substantial quantities of labeled data. Acquiring such data volumes can be difficult, particularly when dealing with faults in machines or structures.
Approach
The proposed approach centers on designing vibration-based IFDS using DTL in scenarios marked by severe data scarcity. A key element is a periodic multi-excitation level procedure, which exploits the inherent non-linear characteristics of real-world systems. This procedure is used to generate images that are then processed by pre-trained Convolutional Neural Networks (CNNs) for fault diagnosis.
The method incorporates a new data visualization technique along with a corresponding augmentation technique. These are designed to mitigate the typical lack of data encountered during the development of IFDS.
Findings
- The approach facilitates the creation of vibration-based IFDS even with limited data.
- A periodic multi-excitation level procedure, exploiting system non-linearity, can generate analyzable images for CNNs.
- The integration of a new data visualization method and its augmentation technique addresses data scarcity.
- Experimental validation on a railway pantograph structure indicated effective support for the proposed methodology.
Why This Matters
The development of methods enabling Intelligent Fault Diagnosis Systems to operate effectively with limited labeled data addresses a significant practical challenge in machine and structural fault diagnosis. This approach could facilitate the deployment of advanced diagnostic tools in scenarios where extensive data collection is impractical or costly, making fault detection more accessible.