A Study on Using Transfer Learning to Utilize Information From Similar Systems for Data-Driven Condition Diagnosis and Prognosis
Prognostics and health management (PHM) is a field of study that aims to diagnose the health of engineering systems and predict their future degradation. In PHM, data-driven methods are among the most popular and empirically validated approaches because they do not require a deep physical understand...
Saved in:
| Main Authors: | , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/11023250/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Prognostics and health management (PHM) is a field of study that aims to diagnose the health of engineering systems and predict their future degradation. In PHM, data-driven methods are among the most popular and empirically validated approaches because they do not require a deep physical understanding of the system properties and degradation behavior. However, training these models typically requires substantial data sets, which are often limited in industry. Additionally, prediction performance is known to degrade when operating and environmental conditions deviate from the training data, or when similar systems with different technical characteristics are considered. Transfer learning aims to address these issues by enabling the transfer of information. The present study analyzes the effectiveness of transfer learning on similar system applications. A rolling bearing and a filter degradation data set are used to evaluate the diagnostic and prognostic performance. The former includes condition data of rolling bearings of different dimensioning, recorded under different operating conditions, and the latter includes degradation data of filters with different filtration areas. Two transfer learning approaches are analyzed: parameter transfer with fine-tuning and retraining, and feature alignment. Both concepts are implemented with the neural network types multilayer perceptron, 1D convolutional neural network, and temporal convolutional network. The analysis demonstrates that current transfer learning approaches can successfully transfer information between similar systems, with improvements exceeding 20 % in some scenarios. However, significant variations in performance are also observed. In certain scenarios, only marginal improvements or even deteriorations are exhibited, revealing prevailing challenges confronting these approaches. |
|---|---|
| ISSN: | 2169-3536 |