Time-frequency method for bearing fault diagnosis

Abstract

Spindle bearings are some of the most critical and vulnerable components in rotating machines. Friction, load forces and vibrations actuating over bearings can produce wear, fatigue and impending cracks on these which may end in a full damage of the spindle over time. The Condition-Based Maintenance (CBM) have arose as a strategy to address this problem, in which, analysis of vibration signals can be performed in real time to anticipate the damage of the machine. A wide range of strategies based on digital processing techniques have been developed for vibration analysis. Wigner-Ville Distribution (WVD) is probably the most used non-linear time-frequency distribution for signal processing in fault diagnosis, however, the presence of cross terms can lead to misleading interpretations of their Time-Frequency Representations (TFR). Signal decomposition methods such as Variational Mode Decomposition (VMD) and Local Mean Decomposition (LMD) have been developed to reduce the complexity of vibration signals allowing to reconstruct them only with their main components. Moreover, this can reduce the cross terms in WVD. However, after the signal decomposition procedure, the identification of the relevant components, which contain the fault information, is commonly based in visual inspection and identification of the bearing housing resonance band. A methodology which combines the great characteristics of the VMD and the WVD is proposed to get more reliable and illustrative results of bearing fault diagnosis from TFR of the vibration signals. Kullback-Leibler Divergence (KLD) was included in the analysis to guide the selection of the effective components with the most relevant information about the fault in an automatic way. After applying the proposed method, in some cases, the amplitude of the fault frequencies in the spectrum were increased around 53% for Outer Race (OR) signals, 45% for Inner Race (IR) signals and 73% for Rolling Element (RE) signals, regarding the amplitude of the found peaks by using the traditional envelope-FFT method. An automatic fault diagnosis method based on an Artificial Neural Network (ANN) and WVD was also presented to avoid the visual inspection. The LMD was used as the signal decomposition method. The TFR, obtained by computing the WVD over the effective Product Functions (PF), were used to build the feature vectors. A classification accuracy in average = 98.2% was obtained by testing the proposed methodology with experimental data.