Incremental rotating machinery fault diagnosis method based on multi-scale knowledge distillation and label smoothing

Abstract Rotary machinery is prone to failures due to its complex and harsh operating environment. Intelligent fault diagnosis methods powered by deep learning have been widely adopted, showing satisfactory performance. However, many methods' applicability is limited to single-task learning scenarios. Mechanical systems typically contain multiple critical components requiring diagnosis. Failure data from different components are collected at various times for model training, essentially forming a task incremental learning scenario. This paper introduces an incremental rotary machinery fault diagnosis system based on Multi-Scale Knowledge Distillation and Label Smoothing (MSKD-LS) to mitigate catastrophic forgetting during incremental learning. MSKD-LS employs a multi-head one-dimensional convolutional neural network as its core framework, leveraging knowledge distillation at directional and distance scales for model knowledge preservation and transfer, and softens real labels through label smoothing to reduce model confidence, enabling significant mitigation of catastrophic forgetting in the absence of replay during incremental phases. MSKD-LS demonstrates effective incremental cross-component fault diagnosis capability in simulated complex mechanical systems with three key components, achieving diagnostic results of 92.92%, 94.33%, and 99.17% across the three tasks. Experimental results show that MSKD-LS can effectively perform incremental cross-component rotating machinery fault diagnosis.