Multilevel Refinement Convolutional Dictionary Learning for High-Speed Bearing Fault Diagnosis

Dictionary learning is widely used in fault diagnosis due to its strong representational capability. However, weak high-speed bearing fault signals are dense in the time domain and sparse in the frequency domain, making it difficult for classical dictionary learning to effectively represent them. This article proposes a multilevel refinement convolutional dictionary learning (MRCDL) for high-speed bearing fault diagnosis. We first design a convolutional dictionary learning (CDL) algorithm with circularly shifted kernels and ridge regression constraints to capture repetitive fault patterns in high-speed bearing signals. To reduce noise interference in the extracted repetitive fault patterns, the CDL algorithm is extended to MRCDL to jointly extract shared high-frequency fault patterns. The proposed method effectively reduces noise interference while also enhancing the weak fault features. Simulations and experimental results demonstrate the performance of the MRCDL.