Incremental learning based intelligent fault diagnosis of rolling bearings

Deep Learning (DL) algorithms excel in identifying bearing failures, but their dependence on large amounts of high-quality data renders them unsuitable for many real-world industrial applications. Additionally, as data is continuously obtained, it can change or new failure types can emerge over time. Consequently, models capable of Incremental Learning (IL) are required to adapt to evolving industrial settings. To address these challenges, this paper integrates Convolutional Neural Networks (CNNs) with Particle Swarm Optimization (PSO) for Intelligent Failure Diagnosis (IFD) of bearings. Furthermore, an IL method is incorporated to mitigate the issue of catastrophic forgetting in streaming data scenarios. The IFD model lever ages PSO to optimize the CNN structure rather than obtaining optimal weights, thereby achieving superior performance in fault recognition. The integrated IL method employs knowledge distillation and replay techniques to minimize forgetting as data gradually includes new failure classes. Various IL algorithms are evaluated under simulated real-world conditions, demonstrating a reduction in catastrophic forgetting. The proposed method proves to efficiently adapt to changes and learn new failure modes, outperforming existing methodologies while reducing computational time.