Lifelong Learning for Bearing Fault Diagnosis with Incremental Fault Types

Abstract

Deep learning (DL)-based fault diagnosis methods have to collect sufficient data on the most comprehensive fault types to ensure reliability. However, in real scenarios, new fault types will be inevitably generated due to complex working conditions, which are also called incremental fault types. Therefore, collecting data for each fault types in advance is difficult. Limited by the characteristics of DL-based models, existing models learning new fault types directly can lead to catastrophic forgetting of old ones, while the cost of collecting data on all known fault types to retrain the models is excessively high. To solve the problem of bearing fault diagnosis with incremental fault types, a new lifelong learning-based diagnosis method (LLDM) is proposed in this paper under the lifelong learning paradigm. The catastrophic forgetting and stability–plasticity dilemma are intrinsic issues in lifelong learning. The core of LLDM is combining the proposed dual-branch aggregation residual network (DARN) with reserved exemplars to overcome catastrophic forgetting and address stability–plasticity dilemma. Two types of residual blocks are created in each block layer of DARN: steady and dynamic blocks, and the adaptive aggregation weights are adopted to balance stability and plasticity. A bi-level optimization program is used to optimize aggregation weights and model parameters. LLDM is applied to a bearings diagnosis case with incremental fault types. Results demonstrate that LLDM is superior to other lifelong learning methods and has satisfactory robustness.