2025/01/01 by Chunlin Li, Qintai Hu, Jianbin Xiong +1
Engineering · #Machine Fault Diagnosis Techniques #Engineering Diagnostics and Reliability #Fault Detection and Control Systems
paper · doi:10.1109/tim.2025.3529071
Accurately implementing compound fault diagnosis for bearings is of great significance. This article proposes an optimized feature entropy extraction method and a novel recognition network for compound fault diagnosis in bearings. First, to address the high noise issue in intrinsic mode functions (IMFs) with high correlation during modal decomposition, an improved ensemble empirical mode decomposition (I-EEMD) method is proposed. This method selects, filters, and recombines the decomposed IMFs, and then extracts multiscale modal entropy (MSME) from the recombined IMFs. Subsequently, a dual-channel multiscale 1-D convolutional neural network (DCMS1DCNN) model is proposed, based on multiscale blocks (MS-blocks) and the concept of residual networks, with convolution kernels determined by a simple and universal rule. The model is designed to recognize MSMEs extracted from mechanical fault signals. By adaptively adjusting convolution kernel parameters to cover a broader range of receptive field (RF) sizes, DCMS1DCNN can utilize deeper features extracted from MSME for fault classification and localization. Finally, the proposed method is validated using a compound fault dataset of bearings provided by the Key Laboratory of Guangdong Petrochemical Institute, along with multiple single-fault datasets obtained from various literature sources. The experimental evaluations conducted on these datasets demonstrate that the method exhibits excellent diagnostic performance for both compound faults and single faults.