Daoming She1, Mingwei Zhou1, Yongyang Wang1, and Michael Pecht2
1School of Mechanical Engineering, Jiangsu University, Zhenjiang, China
2Center for Advanced Life Cycle Engineering, University of Maryland, College Park, MD, USA
For more information about this article and related research, please contact Prof. Michael Pecht.
Abstract:
As a key component of the rotating machinery, the gearbox inevitably develops new fault types during long-term operation. Incremental fault diagnosis can continuously assimilate the knowledge of new faults from streaming data, thereby expanding the model’s capabilities and mitigating catastrophic forgetting. However, in multi-phase incremental settings, the models are prone to the stability-plasticity dilemma. To address the above problem, we present an incremental fault diagnosis method based on a channel-plasticity and dual-memory replay (CPDMR) framework. This method enables continual diagnosis of emerging faults without retraining on the full historical dataset. First, we construct a channel-plasticity convolutional network to evaluate channel effectiveness from channel activation and weight magnitudes. Inefficient channels are then selectively reinitialized to better explore the parameter space and maintain the model’s long-term plasticity. Second, we design a dual-memory replay strategy. A long-term memory buffer preserves important historical information via representative samples. In contrast, a short-term memory buffer organized in a first-in-first-out manner optimizes learning in the current phase. The standard classification loss, a knowledge-distillation loss, and a label-smoothing loss are incorporated to achieve stable and efficient incremental fault diagnosis. Finally, the effectiveness of the proposed method is validated on the experimental gearbox data.
This article is available online here and to CALCE Consortium Members for personal review.