Fault Diagnosis of Rotating Machines based on the Enhanced Multi-Scale Convolutional Neural Network Approach

The fault diagnostics of rotating machinery significantly affect the dependability and safety of modern industrial systems. Advanced fault diagnosis techniques have taken over the challenging and uncertain process of human analysis, boosting the effectiveness of fault diagnosis. The accuracy of fault diagnosis is enhanced by employing deep learning models, well-known for their outstanding ability to process complex relationships across multiple layers. This paper introduces a novel approach to analyzing rotating faulty machines using a multi-scale convolutional neural network (MSCNN) model with an attention mechanism layer. Different faults including unbalanced and bearing faults are investigated based on the pre-processed signals, which are processed using the window-sliding and short-time Fourier transform methods with 13 distinct modes. The proposed approach offers a cost-effective solution without compromising reliability by leveraging a subset of three sensors out of six. The multi-scale CNN model, featuring different kernel sizes, simultaneously captures local and global features. Experimental evaluations demonstrate the effectiveness of the proposed approach in fault classification with an accuracy of 99.68%, which shows the feasibility and accuracy of the proposed technique.