Integrating Demodulation Techniques With Multisource Transfer Learning for Intelligent Fault Diagnosis of Bearings

Profound achievements in intelligent fault diagnosis (IFD) of bearings with transfer learning (TL) based on signal-processing embedded. However, the low performance of diagnosis equipment, large differences in real-time data distribution, and scarcity of fault samples are the main challenges in engineering practice. Most studies have not established the correlation between the signal and the network to enhance the generalization and robustness of the model. This article combines signal demodulation techniques with multisource TL to tackle complex science and engineering problems. First, bearing signal-preprocessing algorithms based on two demodulation methods were developed to construct a bridge between fault feature frequencies and network input sizes. Second, a multisource TL network is constructed to extract bearing fault features and classify health status. Third, a new optimization objective function integrating two alignment methods is designed to decrease the distribution distance between two domains. Lastly, two scenarios are employed to validate the effectiveness of the proposed method which encompasses multisource, cross-domain, and unsupervised TL between the same and different datasets. The result shows that the performance of TL from multisource to single-target is better than TL from multisource to multitarget. In the meanwhile, noise interferes greatly with TL performance, with accuracy increasing as the noise intensity decreases.