Multilevel feature fusion of multi-domain vibration signals for bearing fault diagnosis

Bearing is the key component of rotating machinery, so the rapid and accurate fault diagnosis of bearing is of great significance. As one of the most commonly used diagnostic ways, vibration analysis has been explored by many scholars. However, vibration signals are underutilized in most studies and the fault information is not sufficiently extracted, which will lead to the failure of achieving expected diagnostic accuracy. This paper proposes a multilevel feature fusion of multi-domain vibration signals method for bearing fault diagnosis. Vibration signals are converted into time domain, frequency domain and time–frequency domain for fault diagnosis to realize the full use of vibration signals. The bilinear model and multi-head attention are applied to the fine-grained fusion of features extracted from multi-domain vibration signals. Experiments are conducted on Paderborn bearing data set to verify the effectiveness of proposed method. Results show that the accuracy of proposed method is greatly improved, which is much higher than the other methods.