A novel approach for rotating machinery fault diagnosis using multiscale feature extraction and noise suppression

Abstract In recent years, the research on fault diagnosis methods of rotating machinery has made remarkable progress. However, in practical industrial settings, the efficacy of these algorithms is frequently compromised by varying levels of noise, leading to diminished diagnostic accuracy. This research proposes a novel approach, MDnCNN-MBNet, which integrates a multi-scale denoising convolutional neural network (MDnCNN) with a multi-branch convolutional neural network (MBNet) to address this issue. Firstly, MDnCNN significantly enhances noise suppression capability and multi-scale feature extraction effectiveness by introducing multi-scale perception and attention mechanisms. Secondly, MBNet employs a multi-branch architecture to integrate feature information across multiple receptive fields, thereby enhancing the model’s capacity to represent complex signals. In addition, the wavelet packet log energy maps generated using three wavelet basis functions (db2, db8, and db16) provide rich time-frequency feature inputs for the model. The experimental verification on three datasets shows that the accuracy of MDnCNN-MBNet proposed in this paper reaches 97.65%, 95.61%, and 97.54% respectively in the −10 dB to 6 dB noise environment. Moreover, this model outperforms several advanced models in terms of accuracy and robustness. Consequently, this method provides a promising solution for fault diagnosis of rotating machinery in complex noise environments.