Abstract
In industrial IOT environments, intelligent systems must continually learn from dynamic data streams to maintain diagnostic accuracy for emerging fault types. However, mainstream replay-based incremental learning methods typically depend on knowledge distillation with cumbersome teacher models to preserve past knowledge, resulting in excessive memory overheads that preclude their deployment on resource-constrained edge devices. To overcome this challenge, we propose ME-ACR, a memory-efficient framework driven by attributional consistency reinforcement. ME-ACR first employs cosine-normalized classifier to mitigate the inherent recency bias and establish a stable feature space. Then, two synergistic self-supervised paradigms are designed, which anchors old knowledge without a teacher model. Attributional consistency reinforcement paradigm enforces internal cohesion by promoting consistency among the attribution patterns of correctly classified exemplars. Prototype-distance-based attribution disentanglement paradigm ensures external separation by aligning the topological structure of attribution space with the global geometry of feature space. By preserving the structure of both intra-class and inter-class attributional knowledge, ME-ACR effectively prevents catastrophic forgetting. Extensive experimental results on multiple datasets demonstrate the superiority of the proposed ME-ACR in achieving robust and memory-efficient fault diagnosis under incremental learning scenarios.
Funding Information
  • Scientific Research Foundation for High-level Talents of Anhui University of Science and Technology (2025yjrc0031)
  • National Natural Science Foundation of China (52475116)
  • National Natural Science Foundation of China (62476006)
  • Jiangsu Qinglan Project
  • Fundamental Research Funds of the AUST (2024JBZD0007)

This publication has 25 references indexed in Scilit: