Intelligent Fault Diagnosis of Rolling Bearing based on Incremental Learning

Fault diagnosis in rolling bearings presents considerable challenges due to the dynamic and complex nature of machine operations. In real-world scenarios, systems continually accumulate new data, often including previously unknown fault types. Traditional diagnostic methods frequently struggle to adapt to these evolving fault patterns, making it difficult to maintain accuracy and reliability in fault diagnosis. To address these challenges, this paper proposes the Variational Mode Decomposition Enhanced Gradient Episodic Memory Model (VEGEM). Firstly, Variational Mode Decomposition (VMD) is utilized for data preprocessing, which effectively decomposes signals into key frequency components, isolating essential features and reducing noise. Secondly, the Wide Deep Convolutional Neural Network (WDCNN) is employed as the baseline framework for fault diagnosis, leveraging its deep learning capabilities to handle vibration signals with high precision. Thirdly, the Incremental Learning framework is enhanced using Gradient Episodic Memory (GEM), a constrained replay technique that manages gradient updates to prevent catastrophic forgetting. This method ensures new learning does not disrupt previously acquired knowledge. The proposed VEGEM model was rigorously tested on the CWRU dataset, demonstrating great robustness and adaptability. It effectively incorporated new fault categories and maintained an accuracy rate above $\mathbf{90\%}$ across five Incremental Learning phases. These results confirm that the VEGEM model can effectively resolve the issues of catastrophic forgetting in fault diagnosis, providing a reliable, efficient, and adaptable solution for the Incremental Learning requirements of modern industrial environments.