Vibration Normalization Processing for Fault Diagnostics Under Varying Conditions

Mechanicel vibration signal under varying conditions shows time-varying characteristics, which results in the incomparability of vibration features that corresponding to different machine states, also increases the difficulties in achieving fault diagnosis. In this paper, a processing method called normalization which is capable of relieving the influence of varying conditions is proposed to enhance the vibration-based feature extraction and diagnostics. The normalization processing mainly focuses on vibration amplitude which is quite sensitive to the changes of speed, load and so on. The amplitude normalized signal is derived based on the amplitude envelope of the raw vibration signal. Additionally, a fault diagnosis framework is also proposed by combining the frequency analysis, t-distributed stochastic neighbor embedding (t-SNE) dimension reduction and least square support vector machine (LSSVM) model. The proposed method was verified using the bearing vibration signals under varying conditions, and the analysis results suggest that the normalization processing is beneficial to reduce the feature difference caused by varying conditions and the fault diagnosis method can achieve high accuracy in identifying the bearing state.