Simultaneous estimation of fault type and severity in rolling bearings via multichannel measurement data fusion and a physics-informed measurement model

Abstract

Existing research on rolling bearing condition monitoring has predominantly focused on identifying fault types, with significantly less attention given to quantifying different levels of fault severity. To tackle these challenges, this study introduces a physics-informed YOLO v8 (PI-YOLO v8) model for simultaneous estimation of bearing fault type and severity through two novel physics-informed constraint terms embedded in the loss function. The proposed framework is based on the YOLO v8 architecture, enhanced by incorporating two physics-informed constraint terms into the standard loss function. The first is a fault type classification constraint, designed from the vibration pattern characteristics of different faults, which compels the network to learn tri-axial energy distributions that align with physical principles. The second is a severity constraint, established on the fundamental measurement principle of the monotonically increasing relationship between defect size and vibration amplitude. This term identifies severity levels through a triple-layer constraint mechanism comprising ordinal boundary, consistency, and smoothness regularization constraints, thereby incorporating prior physical knowledge into the measurement process. Additionally, a novel Short-time Fourier Transform-based Signal-to-RGB Image Mapping (STFT-STRIM) method is proposed for fusing multi-sensor measurement data. This technique converts the X, Y, and Z-axis vibration signals into time–frequency images via STFT and maps them to the red, green, and blue channels of an RGB image. Under small-sample training conditions, which simulate challenging measurement scenarios with limited data, PI-YOLO v8 model achieved high estimation accuracies of 97.1 % and 92.9 % for two public datasets, representing substantial improvements of 10.8 % and 11.9 % over the baseline YOLO v8 model.