A Dynamic Voiceprint Fusion Mechanism With Multispectrum for Noncontact Bearing Fault Diagnosis

Abstract

The existing bearing fault diagnosis methods are generally designed to identify the fault type by analyzing the collected vibration data, which belongs to the contact fault diagnosis technology. However, vibration sensors are not only complicated to install but also tend to rub against the equipment during prolonged use, affecting the accuracy of fault diagnosis. To address these issues, a noncontact fault diagnosis method, i.e., dynamic voiceprint fusion mechanism with multispectrum (DVFMMS) is proposed in this article. First, a three-channel feature extractor (TCFE) is designed to extract three different voiceprint features from the raw voiceprint data. Second, an attention mechanism, i.e., multichannel feature fusion self-attention (MCFFA), is proposed to adaptively adjust the network weights of different voiceprint features and realize the dynamic fusion of three voicing features. Finally, the fused features are fed into the classifier to achieve fault diagnosis. In this article, a voiceprint acquisition system based on the VS1053 chip is designed to collect the voiceprint of the bearing operation to accomplish the task of bearing fault diagnosis. The dataset of this study is obtained from the three-phase asynchronous motor platform of Zhejiang University of Technology. The experimental results demonstrate that DVFMMS still performs well under the limited sample size and overcomes the limitations of existing methods.