학술논문

A novel multi-task learning framework with fault mode feature separation for remaining useful life estimation of mechanical systems
Document Type
Article
Source
In Advanced Engineering Informatics March 2025 64
Subject
Language
ISSN
1474-0346
Abstract
With the development of sensing technology and computer science, data-driven remaining useful life (RUL) estimation methods have been increasingly applied in the intelligent operation and maintenance of mechanical systems. However, most approaches use single-task models to achieve RUL estimation under multiple fault modes, which makes it difficult to fully exploit the fault mode (FM) information and its potential correlations with the health state to improve the estimation accuracy. Therefore, we propose a novel multi-task learning framework with FM feature separation capability. First, an unsupervised autoencoder is implemented to cluster the degraded FMs of mechanical systems and label the training samples. Then, a new model named Self-Attention Capsule Network (SA-CapsNet) is established, which utilizes the multi-head self-attention and capsule routing mechanism for degraded feature extraction and fusion, and builds FM identification and RUL estimation sub-networks for multi-task parallel learning. This model has FM feature separation capability of describing different FMs with distinct degradation feature capsules. Furthermore, the RUL estimation subtask employs the masking operation to focus on specific FM degradation feature capsules with the help of FM identification results and utilizes only these capsules for accurate RUL prediction. Finally, experimental case studies on the public turbofan engine degradation dataset and run-to-failure bearing dataset demonstrate the effectiveness and advancement of the proposed framework.