Fault Prognostics for Machinery with Multi-Degradation Stage Based on Deep Residual Network and Multi-Task Learning
- Resource Type
- Conference
- Authors
- Wang, Mingxian; Cui, Langfu; Jiang, Jiahe; Wang, Junle; Jin, Yang; Xiang, Gang; Yang, Jianbing; Lin, Ruishi
- Source
- 2023 CAA Symposium on Fault Detection, Supervision and Safety for Technical Processes (SAFEPROCESS) Fault Detection, Supervision and Safety for Technical Processes (SAFEPROCESS), 2023 CAA Symposium on. :1-5 Sep, 2023
- Subject
- Aerospace
Power, Energy and Industry Applications
Robotics and Control Systems
Signal Processing and Analysis
Transportation
Degradation
Gears
Fault detection
Poles and towers
Multitasking
Feature extraction
Safety
fault prognostics
multi-degradation stage
fuzzy clustering
multi-task learning
deep residual network
- Language
Fault prognostics is critical for complex machines. Deep learning has achieved success in fault prognostics. However, existing methods do not consider the following factors: 1) Machines always take through multi-degradation stages and do not begin to degrade during health stage. 2) Fault prognostics is composed of remaining useful life (RUL) prediction, health stage (HS) prediction and fault category prediction, but previous researches focus more on RUL prediction. This paper proposed a novel prognostics framework to predict both HS and RUL for machines through multi-task learning (MTL). Fuzzy clustering is used for HS labeling and segmented RUL representation. MTL based on hard parameter sharing is used to learn from sensor data. Deep residual network (ResNet) is employed to extract degradation feature as shared layers. Different network towers are designed for two prediction tasks based on shared layers. The proposed framework is verified on XJTU-SY Bearing Datasets. The results demonstrate the superiority of the framework.