Data-Driven Fault Diagnosis for Rolling Bearings Based on Machine Learning and Multisensor Information Fusion

The rolling bearing is known as a crucial component in the turbine pump of a typical liquid rocket engine. Bearing faults may lead to catastrophic events, and effective fault diagnosis is vital to guarantee system reliability and safety. In order to identify bearing faults under extreme operational conditions more accurately and robustly, this article proposes two main intelligent fault diagnoses driven by multisensor information fusion, i.e., shallow model-based and deep learning-based. For the former, the frequency-domain features are first obtained by the Hilbert transform and fast Fourier transform from vibration and sound signals. Then, an improved AdaBoost support vector machine multiclassifier is designed, in which an attention mechanism is used to fuse multisensor information. For the latter, multiple signals are converted into images directly by symmetrized dot pattern and then fused in red, green, blue (RGB) color space. A deep convolutional neural networks model with attention mechanism is further presented to classify bearing faults. The proposed method was tested using rolling bearing experimental datasets containing a large number of raw vibration and sound measurement signals in different health states. The results show the diagnostic accuracy based on multisource signals is better than that based on single-source signals for both deep and shallow models and can deliver a high diagnosis accuracy of 100.0% and 98.8%, respectively. In the case of the same input signal, the diagnosis performance of deep mode is slightly superior to that of shallow model.