Fault Classification of Rolling Bearings Based on Multisource Heterogeneous Data

In the diagnosis of rotating machinery faults, traditional methods often fail to consider the correlation and dependence among various grid features of different signal sources in multisource heterogeneous data. A diagnosis method called self-feature fusion in convolutional neural networks with transformer-based data transformation (SCNNs-TBDT) to tackle the challenges in fault diagnosis for rolling bearings using multisource heterogeneous data is proposed. First, by enhancing the transformer network, heterogeneous data are transformed into a unified format. Subsequently, 1-D and 2-D convolutional neural networks (CNNs) extract grid feature information from the data. Finally, a 1-D convolutional network is adopted to merge the multichannel input data internally and then output the ultimate fault diagnosis results. Experimental results demonstrate that this method exhibits significant advantages in improving accuracy compared to other methods and achieving a classification accuracy of 99.8% on the Paderborn bearing dataset.